ABSTRACT - College of Computing

ABSTRACT RIEDL, MARK OWEN. A Computational Model of Navigation in Social Environments. (Under the direction of Robert St. Amant).

It has long been recognized that humans are fundamentally social animals. Virtually all activities that we

engage in involve interaction with others or are influenced by the presence and opinions of others in the

social setting. Navigation, while studied extensively as an individualistic activity, is no exception. Social

navigation is a term coined to refer to the process of seeking social interaction as source for navigational

support. The purpose of this research is to investigate how allowing a navigating agent to interact with

other agents in a social collaboration affects navigation. A computational model of social navigation has

been developed which models navigation as planning where task-related knowledge can be acquired from

three possible sources: the environment, the agent’s memory, or other agents’ memories. The model is

based on several conceptual models of human cognition for navigation and group decision-making and

made computational by adapting it to decision theoretic planning. Two applications of the computational

model are discussed: the creation of new social navigation tools and the evaluation of virtual environments

where social navigation might be used. TRAILGUIDE is a system developed to aide social navigation on

the World Wide Web and was designed using principals developed from the computation model of social

navigation. MUNE is a Multi-User Navigation Environment, similar to a MUD, which can simulate a wide

variety of spatial and semantic navigation environments. Software agents that are built on social navigation

principals or more traditional models of navigation can be developed to navigate through MUNE in order

to observe their performances. Finally, a simulation of social navigation in arbitrary environments was

designed and implemented to test the reasonability of the computational model. The results are analyzed

for interesting patterns and implications are enumerated.

ii

BIOGRAPHY Mark Riedl is a graduate student in the computer science department at NC State University. He earned his

bachelors degree in computer science with a psychology minor from NC State University in 1999. As a

graduate student, Mark has focused his studies on human-computer interaction and the development of

intelligent user interfaces. Recently he has become interested in the ways in which intelligent user

interfaces affect human behavior.

iii

ACKNOWLEDGEMENTS I would like to thank all those who provided me support, both tangible and intangible, during my efforts

towards my degree. I am indebted to my academic advisor, Dr. Robert St. Amant, for providing me with

the opportunity to choose a thesis topic that interested me and to do research that is uniquely my own. I

would also like to thank my parents, Dr. Richard and Theresa Riedl, for encouraging me to pursue my

interests and for all their support, advice, and unconditional love. Finally, I would like to express my

gratitude to my closest friends for their understanding and patience. Especially Jennifer King, who stood

by me and give me encouragement and emotional support in the last few months, when I was down to the

wire.

iv

TABLE OF CONTENTS LIST OF FIGURES ………………………………………………………………………………... v

1. INTRODUCTION …………………………………………………………………………….. 1

2. COGNITIVE BACKGROUND ………………………………………………………………. 3

2.1. Conceptual Models of Navigation ……………………………………………………….. 4

2.2. Social Navigation ………………………………………………………………………... 6

2.2.1. Review of Social Navigation Literature …………………………………………... 7

2.2.2. Review of Social Navigation Systems ……………………………………………. 8

2.3. Human Information Processing ………………………………………………………….. 9

2.4. Models of Organizational Memory ……………………………………………………… 11

3. COMPUTATIONAL BACKGROUND ………………………………………………………. 14

3.1. Computational Models of Navigation …………………………………………………… 14

3.2. Interleaving Planning and Execution …………………………………………………….. 15

3.3. Communication and Planning …………………………………………………………… 17

4. A COMPUTATIONAL MODEL OF SOCIAL NAVIGATION ……………………………... 19

4.1. Building from a General Framework for Navigation ……………………………………. 19

4.2. Social Interaction as Transactive Memory ………………………………………………. 21

4.3. A Classification Framework for Social Media …………………………………………... 22

4.4. Limitations of the Computational Model ………………………………………………... 25

5. PLANNING AND EXECUTION IN SOCIAL NAVIGATION ……………………………... 26

5.1. Extending the Computational Model with Decision Theory …………………………….. 26

5.2. Extending the Computational Model with Congregations ………………………………. 28

6. APPLICATIONS OF THE COMPUTATIONAL MODEL AND CLASSIFICATION FRAMEWORK ………………………………………………………………………………..

30

6.1. The TRAILGUIDE System ……………………………………………………………… 30

6.2. A Testbed for Evaluating Navigability of Social Environments ………………………… 33

7. SIMULATING NAVIGATION IN SOCIAL ENVIRONMENTS …………………………… 37

7.1. Simulation Design ……………………………………………………………………….. 37

7.2. Simulation Procedure ……………………………………………………………………. 39

7.3. Simulation Results ……………………………………………………………………….. 40

7.3.1. Monotonic Observations ………………………………………………………….. 40

7.3.2. Non-monotonic Observations ……………………………………………………... 43

8. CONCLUSIONS ……………………………………………………………………………… 46

9. LIST OF REFERENCES ……………………………………………………………………… 48

v

LIST OF FIGURES 2.1 Norman’s seven stages of action …………………………………………………………… 3

2.2 A conceptual model of action for navigation ………………………………………………. 5

2.3 Spence’s framework for general navigation ………………………………………………... 6

2.4 The human information processing model …………………………………………………. 10

2.5 The structure of organizational memory as information processing ……………………….. 11

2.6 The transactive memory model for a dyad …………………………………………………. 12

3.1 A model for illocutionary actions …………………………………………………………... 18

4.1 Sources of navigational knowledge mapped to Norman’s taxonomy ……………………… 19

4.2 Decision-making in Spence’s framework for general navigation ………………………….. 20

4.3 The extended transactive memory model …………………………………………………... 21

4.4a Incompleteness in message certainty ……………………………………………………….. 23

4.4b Equivocality in message certainty ………………………………………………………….. 23

4.5 Social media mapped onto the classification framework …………………………………... 24

5.1 A sample expected utilities table …………………………………………………………… 27

6.1 A Screenshot of TRAILGUIDE in use with playback in progress …………………………. 32

6.2a An example spatial MUNE world with description ………………………………………... 34

6.2b An example spatial MUNE world with extensions to the description ……………………... 35

6.2c An example semantic MUNE world with description ……………………………………… 35

7.1a Contour graph of distance traveled affected by the probability of choosing an optimal out-link and the probability of choosing a dead-end out-link …………………………………...

40

7.1b Relationship between the time ratio and the number of social interactions ………………... 40

7.2a Contour graph of distance traveled affected by the social media distribution and environmental uncertainty …………………………………………………………………..

41

7.2b Contour of the distance ratio affected by the social media distribution and environmental uncertainty …………………………………………………………………………………..

41

7.3 Contour graph of the time ratio affected by perceived cost and actual transaction time …… 42

7.4 Contour graph of the distance ratio affected by plan steps and perceived cost …………….. 43

7.5a Contour graph of the distance ratio affected by plan length and social media distribution when perceived cost is low ………………………………………………………………….

44

7.5b Contour graph of the distance ratio affected by plan length and social media distribution when perceived cost is medial ………………………………………………………………

44

7.5c Contour graph of the distance ratio affected by plan length and social media distribution when perceived cost is high …………………………………………………………………

44

1. INTRODUCTION

It has long been recognized that humans are fundamentally social animals. Virtually all activities that we

engage in involve interaction with others or are influenced by the presence and opinions of others in the

social setting [19]. Navigation, while studied extensively as an individualistic activity, is no exception.

Navigation, in its most general sense, refers to the activity of following a route through an environment.

An environment can be any domain in which one has a sense of location and locomotion and is not

restricted to spatial or physical domains. In computer science, navigation is studied by artificial

intelligence researchers, who wish to build robots that can reliably progress through an unpredictable and

changing environment, and by researchers in human-computer interaction (HCI), who use navigation as a

metaphor for transversal of information spaces such as the World Wide Web. Robotic navigation typically

uses some form of planning in which the robot relies on its knowledge of the environment or its perception

of the environment. Information navigation relies on cognitive models of decision-making and

visualization.

Navigation can be understood as situated action [29] where an agent is embedded in the surrounding

environment. The agent can perceive the environment, which affords certain actions such as locomotion.

The agent can also act upon knowledge that is stored in a cognitive map, which is a spatial description of

the environment through which navigation is to occur. A cognitive map is acquired through observations

of the environment over time and is used to find a route to the goal [30, 28]. Having a mental

representation of the environment, such as a cognitive map, is desirable because it allows the agent to look

ahead at possible future states rather than act only on local information [13]. This viewpoint is consistent

with what we know about human information processing: the current environmental situation can be

correlated with the cognitive map, stored in long-term memory, revealing a path through the environment

[30]. In the absence of a cognitive map, the navigational agent could acquire one through direct

exploration of the environment [30, 13], or one could acquire one through the interaction with other agents

situated in the environment. A study of route finding through unfamiliar city roads indicated that a vast

majority of people chose to interact socially with others in order to acquire route-finding information [24,

27]. Other studies support the concept of social communication as a preferred form of knowledge

acquisition in the face of uncertainty [40]. Navigation is a social, and often times collaborative, task [28].

Reasonably, if another person has already expended the effort to navigate a path to a common goal, then

access, through interpersonal communication, to that person’s past experiences can lead to drastically more

efficient path planning [38, 49].

The term, social navigation, has been coined to refer to such a process of seeking social interaction as a

source of navigational support [20, 21, 16, 18]. While there is some debate as to whether social navigation

is distinct from navigation in general social navigation can be distinguished by the tools used to solve the

navigation task [11, 46]. Social navigation is characterized by the use of other people’s experiences in

order to acquire knowledge for navigation [38] in addition to affordances for action that can be perceived in

2

the environment. These experiences are acquired through social interaction with others through a variety of

social media, including computer-mediated communication systems, and are integrated into the one’s own

cognitive map.

The facility for social interaction has an effect on navigation. This paper will review the conceptual and

computational foundations of navigation, both in the physical world as well as in virtual worlds and

information spaces, in an attempt to better understand how social interaction can be integrated into existing

models of navigation. The discussion begins with conceptual models of navigation, the cognitive processes

behind navigational decision-making, and cognitive models of social interaction. Computational

approaches to describing navigation and social interaction will then be presented in the context of planning

and artificial intelligence. Finally, a model of social navigation – navigation in social environments – will

be presented as an extension to existing models of general navigation and described in computational

terms. The description of the computational model will be followed by applications of the model and by a

simulation of navigation that will address the implications of social interaction, as defined in the

computational model, on navigation.

3

2. COGNITIVE BACKGROUND

Navigation has well-founded roots in cognitive psychology, particularly the fields of situated action,

knowledge acquisition, and decision-making. Navigation researchers often assume that the agent

performing a navigation task, be it robotic, human, or otherwise, is situated in an uncertain environment.

Sensory information obtained from the environment may not be accurate or may be misleading or

incomprehensible. Additionally, any action that a robot takes may not deliver it to the desired state since

motor behavior is not perfectly certain. The agent must perceive the environment, refine its understanding

of the environment, and choose an action that will best achieve the goal.

Navigation can have an explicit or an implicit goal. An explicit goal defines a single, well-understood

destination. When the goal is explicit the node identity defines the criterion for success. An example of an

explicit navigational goal is “go to the mall.” Success is defined by being at the designated location and

there is a clear distinction between success and failure. An implicit goal is a set of properties that combine

to define the success criterion. An example of an implicit navigational goal is “find information on the

Web about spider lilies.” What defines success is not so clear because there are many possible solutions of

differing qualities.

Execution of the action sequence

Perceiving the state of the world

Interpreting the perception

Evaluation of interpretations

Sequence of actions

Intention to act

Goals

The World

Figure 2.1. Norman’s seven stages of action [35]

4

Norman lists seven stages of action: forming the goal, forming the intention, specifying the action,

executing the action, perceiving the state of the world, interpreting the state of the world, and evaluating the

outcome [35]. Figure 2.1 shows the seven stages in relationship to the environment. These stages lead an

agent through the process of defining the goal to interpreting whether the outcome of the action was

successful.

Through this series of stages, errors can be introduced in two places: execution of the action and evaluation

of the outcome. Norman defines the gulf of execution to refer to difficulties that the agent has successfully

carrying out the actions required to reach the goal. The gulf of evaluation refers to difficulties the agent has

determining whether it has met the goal. When we consider navigation as action, the gulf of execution

defines difficulties making the correct movements from one location node to the next. This is due to the

uncertainty of motor execution; we assume that correct execution of motor operations will fail some

percentage of the time. During navigation, the gulf of evaluation defines the agent’s inability to determine

whether it is making progress. This is due to the uncertainty of sensation; what the agent perceives may not

always be accurate.

2.1. Conceptual Models of Navigation

Norman’s model of action is generic and applicable to any situated action. The model of action can be

specifically specialized to navigation. One such model of action [29] that has been applied to navigation is

shown in Figure 2.2. This particular model adapts to navigation well. Under the framework for action,

three precursory stages occur before actual, iterative execution begins. First, the agent must establish the

criteria for successful action. A strategy, plan, or policy is chosen that will allow the agent to achieve the

goal. Finally any information the agent requires to execute the strategy is gathered. Iterative execution of

the strategy begins with scanning the environment for information relevant to the task. An assessment is

made as to whether the state of the world is as expected and whether adequate progress is being made.

Based on the assessment of the situation by the agent, environment scanned in the environment may be

incorporated into a conceptual model of the task or a direct action may be made. The action can be a

physical action, changing the state of the world around the agent, requiring the agent to rescan the

environment and begin the executive processes anew. The action can also be a cognitive action such as

selecting new goals, revising the strategy, plan, or policy being followed, or deciding to seek external

information again. As action, navigation fits nicely into this framework [29, 11]. The goal is to find a

particular place or to move to a place where information can be found. A strategy for finding that location

is chosen. Examples of strategies are to use a search engine when navigating online or to follow road signs

when navigating highways [29]. During the acquire data stage the information is acquired from sources

pertinent to the navigation act such as maps. Execution of the navigation act comes in the scan, assess,

model, and act stages. During scanning, one seeks out relevant information about where to navigate to

next.

5

Navigation fits the general action framework well because navigation is situated, requiring close interaction

between the agent and the environment. The local environment surrounding the agent strongly dictates

what navigational acts are possible and where the agent can navigate to next. As situated action, navigation

employs situation-specific knowledge and incomplete information. This reliance on situation-specific

knowledge is referred to as reactive navigation. Another strategy, plan-based navigation, relies on

completely devising a plan – as a series of navigational acts – that will, when executed, deliver the agent to

its goal. Jul & Furnas note that one may often switch between reactive and plan-based strategies in order to

capitalize on the advantages of both [29]. The two strategies for navigation are evident in the framework

for action. Reactive navigation is represented by the scan-assess-model-act loop because, in every iteration

of the loop, situation-specific knowledge is being gathered directly from the local environment, assessed,

and used to further navigation. Plan-based navigation is represented in the decide strategy and acquire

data stages. Switching between strategies can occur through the links from the act stage back up to decide

strategy stage.

Spence has developed a more comprehensive conceptual model of navigation [45], based on the previous

model of action. Spence’s framework for general navigation contains four stages of processing: browsing,

modeling, interpretation, and revision of browsing strategy of processing. Figure 2.3 shows the framework

for general navigation.

Form Goal

Decide Strategy

Acquire Data

Scan

Form Conceptual

Model

Assess Act

Figure 2.2. A conceptual model of action for navigation [29]

6

During the browsing stage, the navigational agent senses the surrounding environment and registers what is

referred to as the environmental content. The environmental content is what the environment has to offer

the agent in terms of the navigational task. Once the environmental content has been registered, the

modeling stage takes place in which the content is used to build an internalized model of the local

surroundings and to add to the internalized global model of the environments that have been visited and

will be visited. During the interpretation stage, the internalized models, both local and global, are analyzed

to determine whether the current locale meets the criteria for successful navigation and, if not, how much

farther the goal is thought to be. If the goal is explicit, interpretation is trivial, answering the question “Am

I at the prescribed destination?” For implicit navigational goals, the interpretation can be more cognitively

demanding, requiring the agent to determine if the attributes of the local environment meet all the criteria.

Failure to recognize success is a failure of interpretation due to the gulf of evaluation [35]. Finally, during

the revision stage, the internalized model of the environment and the interpretation are used to determine

what the next best step, or series of next best steps, will be for the navigation task. The browsing strategy

is a task-oriented plan for interacting with the environment in order to achieve progress towards the

navigational goal. Executing the browsing strategy causes locomotion, delivering the agent to another

locality where it can begin the iterative process again by browsing the environmental content.

2.2. Social Navigation

The term social navigation refers to the ways in which perceived social factors influence navigational

behavior. Dourish and Chalmers define social navigation as moving towards a cluster of people or

navigating to a particular place because someone else has been there or seen something [20]. The first

definition of social navigation, movement towards a cluster of people, is most often investigated in terms of

collaborative virtual environments (CVEs). Typically a user is immersed in a 3D environment and often

represented by a graphical avatar. Other avatars also inhabit this space, the goal is to locate and interact

Content

Internal Model

Interpretation

Browsing Strategy

Model

Interpret Revise Strategy

Browse

Navigation

Figure 2.3. Spence’s framework for general navigation [45]

7

with the other inhabitants of the virtual world. The second definition of social navigation, navigating to a

place because someone else has been there or seen something, is most commonly addressed with

recommender systems. Both descriptions refer to a form of navigation in which communication with

others, either directly or indirectly, is essential for successfully reaching the goal. Dieberger more broadly

defines social navigation as the sharing of navigational information between people through

communication [16, 18].

Navigation in environments where social interaction is possible differs significantly from asocial

navigation. Navigation is dependent on the source of way-finding information. Asocial navigation

depends solely on sensing the environment and developing conclusions from sensory information. In a

social environment, an agent can choose to receive way-finding information from others through the course

of social information. While social interaction could be considered merely another type of environmental

resource, the number of factors affecting how social communication takes place and how the results are

understood are numerous and complex. Navigation in social interaction becomes less a system of sensing

and acting, and becomes more a decision process on where to turn for way-finding information: the simple

sensing of the environment, or the complex process of social interaction which can be potentially very

rewarding.

2.2.1. Review of Social Navigation Literature

The earliest example of social navigation is presented in Vannevar Bush’s description of the hypothetical

MEMEX system [8]. In MEMEX, Bush imagined academicians would be able to collect huge repositories

of microform documents (the state of the art document storage and retrieval at the time) in a special

machine. These microform documents could be annotated with personal comments and notes, and

furthermore, like documents could be linked together sequentially to create thought trails. The creation of

trails is an explicit action to guide future navigation through a semantic space consisting of documents.

Bush describes how the owner of the MEMEX system can use the trails to guide himself and also how the

trails of microform documents can be passed on from mentor to student so that the students can follow the

contextual trails the mentor relied on. The MEMEX system described by Bush was never realized,

although a modern document archive, the World Wide Web, is reminiscent in many ways.

Dourish and Chalmers [20] were the first to explicitly divide navigation spaces in spatial models, semantic

models, and social models. The spatial navigation model refers to navigation spaces that perpetuate

inherent spatial qualities, such as the physical world, maps, and virtual reality. The semantic model refers

to navigation spaces that are not spatial, but rely on understanding the relationship between information

objects. Examples of semantic spaces are databases and the World Wide Web. The social navigation

model is described as being distinct from the spatial and semantic models because navigation relies on

understanding the actions of collaborators. Just as spatial navigation models can be applied to semantic

spaces by laying out information objects in a spatial format, social navigation can overlay spatial and

8

semantic spaces. In a spatial framework, collaborators have a location in space and social navigation may

involve moving towards them in order to participate. In a semantic framework such as the World Wide

Web, personal Web pages and collaborative filterers recommend courses of navigation based on social

context [20].

Erickson [21] independently reached the same conclusion about personal Web pages. Erickson views

personal Web pages as a strategy for users to publish information and construct a publicly available

identity. Furthermore, personal Web pages often have hyperlinks to other pages. These hyperlinks are

contextual in that they represent people and places that are interesting to the author and act as

recommendations to others. This recommendation functionality of personal Web pages changes the task of

finding something on the Web into a task of finding someone who knows where to go. This is analogous to

the way in which humans socially collaborate in the real world [21] when one asks for directions.

Dieberger [16] further expounds on the use of personal Web pages as tools for social navigation but also

discusses how email and newsgroups have been converted into tools for social navigation. Emails and

newsgroups both provide opportunities for users to enumerate Web URLs in the body of a message as a

way of indicating to others that the document belonging to the URL is in some way significant. A reader of

that message who chooses to enter the URL into their Web browser has effectively undergone social

navigation. Dieberger goes on to list ways in which existing computer-mediate communication media,

such as email, newsgroups, and MUDs, can be designed to make social navigation more natural.

While most social navigation researchers focus their efforts on the World Wide Web because the size and

unstructured nature of the semantic space lends itself well to navigation through collaboration, Riedl [38]

discusses social navigation as a phenomenon that occurs in the physical world, virtual worlds, and in

semantic spaces such as the Web. Riedl introduces a computational model of social navigation that is the

foundation of this thesis. The model is presented in detail in section 4.

2.2.2. Review of Social Navigation Systems

The first systems recognized as tools for social navigation were computer-mediated communication

systems that were adapted to the task of sharing navigational knowledge about the World Wide Web. On

the World Wide Web where the problem of finding useful material is compounded by its size and

unstructured nature input from others through social exchange becomes even more essential. Recognizing

the need to exchange information about Web sites, users published bookmark pages [21, 18, 16].

Bookmark pages were Web documents created by individuals to point to other Web sites of interest. Any

Web user who shared similar interests with the bookmark page author might follow the hyperlinks to other

sites of interest [21, 16]. Similarly, email and newsgroups also became mechanisms for informing others of

relevant documents on the Web [16]. URLs could be embedded in email and newsgroup postings

indicating that those who read the message should browse to that location. Lately, recommender systems

have become essential tools for social navigation and for creating a sense of community on the Web [26].

9

On the Web, the most common recommender systems are collaborative filterers that match a users profile

to aggregated data about where others with similar profiles have browsed [36]. While social exchange is

not direct, as with Email, newsgroups, and bookmark pages, information is shared about what others like to

view and the experiences of social peers are exchanged.

The JUGGLER 2.0 system is a MOO (MUD Object Oriented) that meshes the World Wide Web with a

textual MUD environment [15, 16]. Users in a discussion can speak about URLs and to point out Web

pages as if they were objects in the JUGGLER world. When the JUGGLER system recognizes an URL, it

automatically causes the client user’s Web browser to load that page. JUGGLER users can refer to the

Web this way in order to facilitate navigation of Web pages that are relevant to the discussion [16].

Additionally, URLs can be embedded into the world itself so that moving from one room to the next causes

simultaneous navigation from one Web page to another. Hallways can be created to define a path through

the Web even when the referenced Web pages are not hyperlinked together [15]. By creating such

hallways, users can recommend sets of contextually related Web pages to those who visit in the JUGGLER

MOO.

A recent interest in social navigation and recommending on the World Wide Web has lead to more

specialized systems for sharing navigation experiences. The FOOTPRINTS system [49] records histories

of user navigation on the Web and presents it as an anonymized and aggregated directed graph of particular

Web sites. The FOOTPRINTS system provides digital cues pertaining to usage of digital documents on the

Web. Without such cues, any user must assume that he is the first and only user to navigate to a particular

site [49]. Through FOOTPRINTS, one can clearly see where colleagues have gone in the site and can thus

focus one’s search.

The WALDEN’S PATHS system [23, 42, 43] is a pedagogical tool that allows teachers and peers to

annotate Web pages and link related pages to create paths, similar to those in MEMEX. By creating a path,

a user is explicitly recommending to other users that a certain set of Web documents be visited in an

explicit order, thus preserving context. Like the JUGGLER MOO, the documents in the path do not need

to be hyperlinked together.

2.3. Models of Human Information Processing

Although social navigation is widely observed in human activity, computer agents can also be programmed

to interact socially. Computer agent intercommunication is a challenging field of study in computer

science, whereas intercommunication comes naturally to human beings. Navigation models can be useful

but are incomplete without a cognitive component. Hutchins refers to navigation as the manipulation of

mental symbols that represent some physical representation relevant to navigation [28]. It makes sense,

therefore, to investigate how humans process information and use that information to make decisions.

Human decision-making is often modeled as information processing. Information is taken in through the

senses and, through a series of cognitive transformations, a decision is made about the most appropriate

10

behavior to perform [50]. The model of the human as an information processor is presented in Figure 2.4.

Information is collected from the environment through the senses. Through a combination of bottom-up

and top-down processing, and attentive filters, cues are recognized that are important to the decision-

making task. A diagnosis is performed on the cues by encoding the sensory information into a pattern

through the use of heuristics. The pattern is a key with which to trigger activation of stored memories in

long-term memory. These memories can be thought of as records of how problems were solved in the past.

When stimuli are encoded, they will trigger a set of responses stored from past experiences. The

appropriateness of the response is determined by the closeness of the match between encoded situation and

recorded experiences. Novel experiences may not match very closely with previous experiences at all,

depending on the matching heuristic. The best response is selected and the response action is performed,

somehow altering the surrounding environment (from the perspective of the information processor). Given

the change in the environment, new cues must be perceived in the environment and the process iterates.

The process of perceiving contextual cues, encoding through heuristics, and comparing to long-term

memory is often referred to as situation awareness [50]. The best option is chosen from the possibilities

retrieved from long-term memory. Optimally such a choice can be determined through a decision theoretic

process of analyzing expected costs, probabilities, and values. However, cognitive scientists find it more

likely that humans rely on a series of simple heuristics and biases [50]. Finally the chosen response is

executed, altering the agent’s environment providing new or altered sensory information [50].

Conceptually, long-term memory can be thought of as a repository in which action responses are stored.

Decision-making, likewise, is the process of using perceptual cues to index into this repository and retrieve

one or more possible actions [50]. Of course, the exact pattern of cues used to trigger a response is not

always available in long-term memory. In such a condition, the best match must be found. Compared to

Senses

Environment

Perception Diagnosis

Working Memory & Cognition

Long-term Working Memory

Choice Response

Long-term Memory

Figure 2.4. The human information processing model [50]

Situation Awareness

Options, Risks,

Values, Probabilities

11

the complexity of the human mind, the human information processing model is an extreme simplification

but has been found to be a reasonable approximation of human behavior for the purposes of HCI.

2.4. Models of Organizational Memory

Since social navigation involves collaborative effort among several agents, the human information

processing model of decision-making cannot alone describe how the navigating agent makes decisions in a

social environment. In a similar fashion to modeling an individual as an information processor, one can

also model a group or community as an aggregate information-processing organism [47]. Abstraction of an

organization, consisting of any number of individual members, to a single information-processing entity

allows researchers to fall back on the well-understood models of individual decision-making and apply

them to group decision-making. Within a social organization, the collective memory of that organization,

as a whole, is distributed among the individuals and their artifacts [39, 47]. Figure 2.5 shows the

conceptual structure of an organization as an information-processor. As the community changes the

knowledge collectively known by the group changes, although through the transfer of knowledge, beliefs,

and culture between the members of the organization, much of the organization’s collective memory is

persistent [39, 47].

While it is convenient to aggregate the community members’ individual memories into a group memory,

such a simplification only exposes the issue of social communication. Because a group memory contains

some relevant information, it does not mean that an individual has direct access to that piece of memory.

Knowledge is no longer guaranteed to be a simple process of retrieval when transitioning from the

individual to a group or organization. The issue of locating one who holds the knowledge and of

Organization Information Acquisition

Retention Facilities

Individuals Culture Ecology Structure

Information Retrieval

External Archives

Decision Environment

Organization Information Acquisition

Retention Facilities

Individuals Culture Ecology Structure

Information Retrieval

External Archives

Decision Environment

Figure 2.5. The structure of organizational memory as information processing [47]

12

communicating with that other person to acquire that knowledge replaces the issue of simple memory

recall. That individual may be required to locate which other individual holds that information and to

engage in communicative interaction with the holder in order to retrieve that information. In an

organizational sense, long-term memory retrieval becomes an issue of social interaction [39, 48, 28, 32]

and can be modeled as delegation of information retrieval and decision-making [39, 28].

Wegner [48] formalizes the notion of organizational memory by comparing a group memory to a

distributed network computer. A distributed network computer contains many individual processors, each

containing part of the data. Each processor contains a directory that indicates whether a piece of data is

contained within the local memory store, or a remote processor’s store. Any memory retrieval first

consults this directory before proceeding with the retrieval. If the data is not local, then the specified

remote processor is queried over the network. Wegner likens humans in a community to such a distributed

network computer [48]. Each individual has a local memory store containing part of the group’s shared

knowledge and a directory indicating where any piece of knowledge can be accessed. If a piece of

knowledge is not local, one must contact the other who does know and engage in social communication in

order to transfer the knowledge from the other to oneself [48]. Do to the inherent complexity and varieties

of ways in which social communication can take place between two humans, Wegner formalizes the

communicative process of knowledge transfer as a transaction, aggregating all communication from the

initial request for information exchange to the final completion of transfer as well as all social protocols

invoked. Figure 2.6 shows a schematic of how agents are represented by Wegner’s transactive group

memory.

In the case of human agents, the directories of remote memory stores may not be completely accurate. The

remote directories contain indexes of what one believes others to know. The directory may be built upon

Person A

Local Processor

Local Memory

Local Directory

Remote Directory for

Person B

Person B

Local Processor

Local Memory

Remote Directory for

Person A

Local Directory

Figure 2.6. The transactive memory model for a dyad (adapted from [48])

13

previous interactions and experience socializing with others or may also be built upon stereotypes applied

to others. Do to the impossibility of completely accurate directories, one may often choose an

inappropriate target for memory retrieval and, hence, to interact with [48]. The underlying assumption

behind the transactive memory model is that one chooses local or remote memory sources to transact with

because directory lookups are actually attempts to match stimulus with possible responses stored in long-

term memory. The stimulus is recognition of the need for knowledge. The possible responses are

transactions with sources of knowledge and are chosen by matching the need with those transactions that

will best alleviate the need through a heuristic search.

14

3. COMPUTATIONAL BACKGROUND

Artificial intelligence, a field of study with close ties to cognitive psychology, has taken a different

approach to the study of navigation. Many researchers in the field of artificial intelligence are interested in

building robots that can operate successfully in the world. Practically, robots must be situated in uncertain

environments the same way that humans are and be able to function effectively. As a form of applied

science, researchers strive to build computational models of the behaviors they observe in the real world.

3.1. Computational Models of Navigation

Robotic motor effectors are not perfectly guaranteed to deliver the robot to the specific desired location and

robotic senses often detect the environment incorrectly. Robots must compensate for these uncertainties in

they way they plan their paths through the environment. The most common models used for computational

navigation are models of situated action, hierarchical planning, and stochastic models.

Models of situated action provide designers with a firm technique for dealing with uncertainty in the

environment. The agent is assumed to be situated within a dynamic and uncertain environment and that the

agent can accomplish its goals by means of reactive search. The agent senses the environment, which

affords a specific set of actions that can be performed. The agent chooses the action that will most likely

propel it closer to its goal and then performs that action. These models, in their pure form use no

representational knowledge of the environment; the agent is merely reacting to the environment in a pure

stimulus-response manner [4]. Reactive agents are shown to be the most robust in terms of how they deal

with dynamic and uncertain environments, but since higher level cognitive activity is not yet generally

possible in reactive systems, the agents are limited to mimicking low-level life forms [4] such as insects.

The hierarchical planning model uses abstract representations of the environment to build plans that an

agent can execute in order to achieve a goal. At the highest level, a plan can specify the actions that the

agent must perform without dealing with issues of uncertainty. As the abstract, high-level plan is

decomposed into a more specific plan, the agent focuses its attention on one specific area of the

environment, attempting to resolve any uncertainty there. As the level of planning becomes more and more

specific, rapid real time planning becomes essential. If the agent moves to a state that is unexpected,

replanning is invoked. The agent can have a model of the world to work with before navigation begins, or

can develop the world model as it interacts with the world. The agent is able to understand the scope of the

world as it relates to its navigation and can thus avoid pitfalls and unnecessary actions, but at the expense

of not being able to deal effectively with uncertainty [4]. The use of hierarchical plans, as opposed to serial

plans, allows agents to reason about navigation in a way that is similar to the way humans plan [33].

Typically hierarchical path planners for navigation use a three-layer hierarchy, as is seen in AURA [3].

The top-most layer is a general mission planner that is responsible for providing the lower layers with the

correct parameters and determining the criteria for the navigation. The middle layer is a navigator that

15

plans the path at a general level of detail. An example of a general path plan would be “proceed 25 miles,

turn left at traffic light” [3]. Finally the lowest layer, referred to as the pilot, focuses only on the most

current sub-goal by transforming it into sequences of motor actions that are prone to failure.

The third model is to model the uncertain environment in a stochastic framework. In such a model, the

agent is tasked with finding a path that will succeed with the highest probability of success. The most

common technique for stochastic computation is Markov chains, which is equivalent to a Bayesian network

description of the environment. Markov chains can predict probabilities for chains of actions when the

probabilities of success of chains of actions are known, only partially known (partially observable Markov

chains), or completely unknown (hidden Markov chains). Dean et al. [14] describe an algorithm for

planning using Markov chains. The algorithm, given a directed graph and a reward function, a policy

governing agent behavior is generated. In the event of unsuccessful execution of an action, the policy will

be able to provide the agent with the next best action to perform [14]. Policies, in this sense, are very

similar to the universal plans described by Schoppers [41]. Due to the computational complexity of

Markov chains, partial policies, referred to as envelopes, are generated only for the nearby environment.

As the agent moves through the environment, the envelope is expanded and then pruned. The stochastic

model, however, is completely unrelated to human cognition in that humans cannot compute the

probabilities of their actions to the degree required by this technique.

3.2. Interleaving Planning and Execution

As seen in section 2.1, navigation is a problem-solving behavior closely related to path planning. The

parallels between navigation and planning are clear and unambiguous. As with most artificial intelligence

problem solving, the goal is to transform the agent’s current state – human, software, or robotic – to a

solution state. In the case of navigation, the states in the problem-solving domain are locations in the world

and the goal state is some local environment that meets specified criteria. The agent transitions between

states through locomotion. In navigation, since sensation is often needed to determine the successors to a

given state, state transitions may be costly in terms of the cost of motor actions required to generate

locomotion. Backtracking can become even more costly.

There are many techniques to planning, spread across a continuum from classical planning to purely

reactive planning. Classical planning refers to techniques where all information about an unchanging

domain is known up front and the agent must arrange a set of operators into a sequence that will guarantee

success. The sequence of operators is executed only once the planning is complete. Classical planning for

navigation makes the unlikely assumption that the world is completely known and that motor execution

will never fail to deliver the agent into the desired world state. Purely reactive planning takes the opposite

approach and only chooses the next operator to be executed, iteratively switching between opportunistic

operator selection and execution. Reactive planning for navigation is a commonly used technique because

it handles the uncertainty of realistic environments well. However, a purely reactive agent only relies on

16

the environmental information that it can sense in its current state and is often unable to plan several steps

into the future, making backtracking a frequent necessity. Often an intermediate planning technique,

collectively called interleaved planning and execution, is used.

Interleaved planning and execution recognizes the fact that classical planning can be performed as long as

there is enough information. When conditions arise that necessitate action in order to retrieve further

information, the planner can begin executing the partially constructed plan until information is again

available and planning is again possible in the classical sense.

The data structures used to represent a plan are of crucial consideration when interleaving planning and

execution. Purely reactive planners do not use any plan structure at all since only one operator is ever

considered at a time. STRIPS and STRIPS-like structures are used in classical planning because it captures

ordering constraints between operators, making the construction of sequences of operators easy. However

STRIPS assumes that the world domain is perfectly known so the only constraints needed are ordering

constraints; STRIPS, in its original capacity, cannot handle constraints imposed by lack of information.

The IPEM system [2] used a revised STRIPS planning language to handle incomplete information by

introducing lack of information as a flaw to be corrected. Lack-of-information flaws are repaired through

execution. Etzioni et al. have also built a non-STRIPS based grammar, UWL, to handle planning when

some operations are meant for information gathering so that planning can be more complete [22]. Both

IPEM and UWL operate on the principal that there are two kinds of operations, those that progress the

agent through the domain’s state space, and those that gather information so that planning can be more

robust. Those operators that gather information can only be executed when the agent has reached the state

defined by the operator’s preconditions. For the agent to reach this state, execution of the partial plan must

occur. Planning and execution are interleaved because planning can resume once more information has

been gathered.

In a navigational domain, interleaving planning and execution makes sense. Classical planning assumes

extensive knowledge about the global environment through which the agent will be navigating is known.

This knowledge would have to be extensive enough that the agent could rely on that pre-existing

knowledge to perform all motor behaviors required for locomotion without fear of motor error [4]. No

sensing would be required. The assumption about a static and extensively known world is not valid in a

world in which there is uncertainty. In most cases, foreknowledge cannot be expected to be complete and

the agent will be expected to navigate through regions of which it knows nothing in order to gather more

information. Often, the environment through which the agent is navigating will be so uncertain that the

agent will never be able to plan more than one operation in advance before having to execute in order to

gather more information. For this reason, navigation is considered an iterative cycle of sensation, selection

of the next operation, and execution [29]. When the uncertain environment is the only source of

knowledge, planning will be very reactive.

17

The difficulty of interleaving planning and execution does not lie in either the planning or execution that

the agent performs. Planning can be achieved by the variety of techniques discussed in section 3.1. The

difficulty of interleaving planning and execution is deciding when to switch from planning to execution and

from execution back to planning. This is a problem that virtually every navigational path planner must

address. SHAKEY [34], an early system that used planning to perform navigation, would halt execution

when errors in motor control brought the system to an unrecognized state and re-invoke the planner. In

SHAKEY’s case execution was performed when planning was complete and the planner was re-invoked

only when the plan failed. Similarly, AURA [3], which utilizes hierarchical planning instead of serial

plans, invokes replanning when the lowest levels of motor execution fail. A more sophisticated system,

HUEY [13], uses decision-theory to determine whether planning or execution will yield the highest degree

of information gain. For navigational agents using the stochastic model, we can observe how the method

described in [14] has been used to simulate robotic navigation. The algorithm interleaves planning and

execution by determining when to allocate time to generating policies (planning) and when to expand and

refine the envelope through execution. The interleaving occurs when a trigger event occurs, such as the

expiration of a deadline or the violation of an envelope.

3.3. Communication as Planning

Communication is a form of planning between agents and can be modeled computationally as actions that

change the state of the world in a predictable and desirable fashion. Utterances by an agent change the

mental states of both the speaker and listener. Illocutionary utterances – a request or a question – act in this

capacity to inform the listener that the speaker is in a certain mental state and is intending the listener to

adopt a certain mental state [10]. Illocutionary acts are considered primitive in that they cannot be broken

down into smaller actions and that these primitive illocutions, as requests or questions that are to be

answered, will prompt one or more response events [10]. Response events include performing an action,

changing the state of the world, or answering a question, changing the mental states of the agents involved

in communication. Figure 3.1 shows the model of illocutionary actions.

In planning terms, illocutionary acts are operations that can bring about changes in the state of the world;

an agent will perform an illocutionary act when it desires the post-conditions of the action. Given a certain

precondition, C, the illocutionary action, A, will result in the events, E1, E2, E3, …, and EN, each with

corresponding post-conditions, C1, C2, C3, …, and CN. Furthermore, since illocutionary acts can be

performed in many different ways (e.g. a direct request, an indirect question, word choice, etc.), it is useful

to abstract an utterance away from the phraseology itself and just refer to the mental state of the speaker

and the intended mental state of the listener. When illocution and the events that respond to it are

abstracted to primitive operations, the model of illocutionary action becomes reminiscent of the STRIPS

planning language in that the post-condition of one operation is the precondition for the following operation

in a plan. It makes sense to refer to communication, at least in the sense of illocution, in terms of planning.

18

NCCCC EEEEAC N →→→→→ −1321

321: �

Figure 3.1. A model for illocutionary actions

In social navigation, the navigating agent may have a mental state consisting of an incomplete plan. The

plan is incomplete because the agent is missing information about the environment that it needs in order to

plan a successful path through the environment to its goal. Given this mental state, the agent may choose to

perform an illocutionary action – the action could be a question: “how do I get to…” or a request: “show

me the way to…” – because a response will change the agent’s mental state by filling in the missing

information. In this sense, the agent is sharing its goals with another agent who will adopt those goals as its

own and act upon them [10]. It does not matter how the agent attempts to perform the action or ultimately

which words were chosen, only that the agent had a given mental state before the illocution and desired a

certain mental state after illocution. In planning terms, the illocutionary act can be described completely by

the preconditions and the post-conditions.

19

4. A COMPUTATIONAL MODEL OF SOCIAL NAVIGATION

Existing conceptual models of navigation are not sufficient to deal with the complexities of situating a

navigational agent in a social environment. While at the most general level, navigating in social

environments is still fundamentally a task of sensing, planning, and execution, the possibility of social

interaction as a source of navigational direction introduces the complexities of inter-agent communication.

Traditionally, navigation is viewed as the interaction between the agent and the environment and that all

navigational decisions stem from observations of the environment. Another source of decision-making

knowledge is the agent’s memory; the navigational agent may already possess the knowledge needed to

choose the next locale to navigate to. Social environments yield a third possible source of navigational

knowledge: other agents in the environment who have previous experience. The agent can choose to

interact with other agents socially in order to learn from their experiences and thus delegate the decision-

making process to another agent instead of relying on the environment or personal memory.

Sources of Navigational Knowledge Norman’s Taxonomy

Environment

Social interaction Knowledge in the world

Memory Knowledge in the head

Figure 4.1. Sources of navigational knowledge mapped to Norman’s taxonomy

Norman [35] speaks of problem-solving knowledge as residing in two different places: in the head and in

the world. When knowledge resides in the head, it is in the form of procedures and rules learned in the

past. Knowledge in the world refers to procedural knowledge that can be acquired by directly interacting

with the environment. For example, a map encodes navigational knowledge in the environment so that the

agent does not have to implicitly know how to reach a destination. This information is provided to the

problem-solving agent externally so that it does not have to rely on cognitive resources in the head or rely

on knowledge that may not be present in memory. Other agents who are in the surrounding environment,

or who can be communicated with also represent resources in the world that can be drawn upon instead of

cognitive resources. In terms of navigation is social environments, we can decompose Norman’s notion of

knowledge in the world in to knowledge in the environment and knowledge in others’ heads. Figure 4.1

shows the three sources of navigational knowledge in a social environment mapped onto Norman’s

taxonomy.

4.1. Building from a General Framework for Navigation

Navigation in a social environment fundamentally is no different from navigation in a social vacuum. The

only thing that changes when an agent performs in a social environment is the richness of sources of

20

decision-making knowledge that can be drawn from. To build a computational model of social navigation,

we can therefore start with Spence’s general framework for navigation as discussed in section 2.1.

It is during the revision stage that knowledge about the environment is used to perform decision-making.

As mentioned previously, this navigational knowledge comes from the environment, the agent’s memory,

or others through social interaction. These three sources of information are collected into the internal

model of the environment and it is this model plus the interpretation of the model that is used to revise the

browsing strategy. Knowledge in the head can be interpreted as a pre-existing model of the navigational

task held by the agent before the navigational task begins. Knowledge in the world – both that from the

environmental content and knowledge in others’ heads – is acquired during the browse stage and appended

to the internal model. Therefore, when the strategy revision stage comes, the choice of action is a matter of

choosing the best response, based on heuristic pattern matching with the internal model stored in memory.

Figure 4.2 shows how the stages of Spence’s framework for navigation [45] interact with long-term

memory to form the decision-making process of navigation. They gray region is long-term memory. The

internal model of the environment, which is stored in long-term memory, is interpreted with the use of

heuristics. The internal model plus the interpretation form the basis for choice.

The revision of the browsing strategy itself can be thought of as planning. Knowledge about the current

state of the agent is considered and a decision about the next operations to be performed is made. In the

domain of navigation, the current state is the local environment and operations are those that move the

agent from one locale to the next. During the revision stage, the agent must decide whether it is more

beneficial to stay and attempt to add to the plan or whether there is greater benefit in executing the partial

plan. One motivation for execution before planning is complete is to move the agent into a state where

information is available and further planning is possible. In the worst-case scenario, the agent will execute

Content

Internal Model

Interpretation

Browsing Strategy

Model

Interpret Revise Strategy

Browse

Navigation

Figure 4.2. Decision-making in Spence’s framework for general navigation

Long Term Memory

21

completely to the end of the partial plan and will be in the position to reactively plan the next operation.

For the agent to decide that further planning is beneficial, it must believe that it can extract more

information from its current state. When planning is chosen, the agent must also choose the source: the

environment, its memory, or social interaction.

4.2. Social Interaction as Transactive Memory

During the revise strategy stage, the decision to communicate, as a form of planning, is made. Before now

we have considered the agent to only consider whether to communicate or not. Crucial to communication

as planning, the agent must choose whom to communicate with. We assume that the agent will, from a list

of people with whom communication is possible, choose to communicate with that which is most likely to

respond favorably. Furthermore, since the decision is being made to communicate, it would be desirable to

model this decision in a way that captures the decision to plan based on all three possible sources:

environment, memory, and others.

When the decision to communicate for the intent of planning – the decision to perform an illocutionary act

– is made, planning becomes a group decision-making process. As discussed in section 2.4, identifying the

person who holds the desired knowledge becomes crucial. Wegner’s transactive memory model provides a

good, computational framework for determining whom to interact with. Wegner’s model provides that the

agent has remote directories in memory that index what the agent believes others to know [48]. Deciding

with whom to interact is an issue of querying all directories simultaneously. Directory querying handles

the determination of whom has the information needed for the next operation. Of course, the directories

may indicate that the information is already local to the agent and in the agent’s long-term memory.

The World

Local Agent

Local Memory

Local Directory

Env.

Directory

Remote Directory

Local Processor

Remote Agent

Remote Memory

Remote Processor

Figure 4.3. The extended transactive memory model

22

The transactive memory model already covers two of three sources of knowledge, both knowledge-in-the-

head, and knowledge-in-others-heads. The model can be extended to include knowledge-in-the-

environment by adding environmental directories to the agent’s memory. An environmental directory

indicates what kind of navigational knowledge can be acquired by interacting directly from the

environment at specific locations. For example, a certain location, such as a visitor center, will have a map

that can be examined. The environmental directory would indicate that being in such a location that holds

value, in terms of acquiring navigational data needed to achieve the goal. Figure 4.3 shows the extended

transactive memory model with environmental directories.

Deciding where to acquire planning knowledge is reduced to the process of performing simultaneous

directory queries. Once a source is chosen based on the directory query, any interaction with the source in

order to acquire the knowledge is abstracted as a transaction [48]. The model does not specify how the

transaction takes place, only that the transaction ends when knowledge has been transferred to the local

agent. The transaction is an illocutionary action and, according to the model of illocution proposed by

Cohen and Levesque [10], the wording of the illocutionary act is not as important as the fact that the local

agent’s goals and beliefs are transferred and one or more events will transpire in direct consequence. These

events should encapsulate the transfer of knowledge from the source to the direct agent, completing the

transaction. Of course, it should be noted that one possible source of interaction is the environment and

that illocution with the environment does not make sense. In this special case, one can consider the

transaction with the environment as using the senses to draw information from the environment that

directly addresses the agent’s goals.

The local directory, remote directories, and environmental directory all operate simultaneously to produce

the solution with the lowest perceived cost. The perceived cost of any transaction is comprised of many

factors such as the time to perform the transaction, the time to interact with content in the environment,

perceived cognitive effort for local memory recall, desirability of a specific course of action, uncertainty,

and so forth. Many of these cost factors are determined by individual preferences and cannot be modeled

deterministically. Although there are many ways of measuring cost, our model will consider only the time

to complete the transaction. The time to complete a transaction is dependent on a variety of factors,

including personal performance and media choice. We will assume that the agent will always choose the

transaction with the lowest perceived cost.

4.3. A Classification Framework for Social Media

In our computational model of social navigation, the perceived cost of transacting with a knowledge source

is assumed to be related to the time it takes to complete the transaction through a social medium. Social

medium, in the context of this paper refers to any mechanism through which information can be transferred

among two or more agents. Social media are thus not restricted to computer-mediated communication, as

communication can occur through a wide variety of means, including face-to-face dialog. Regardless of

23

the social medium chosen, a transaction is going to consume a certain amount of time, dependent on the

social medium through which the transaction occurs and with whom the transaction occurs with, there are

also confounding factors such as message certainty. Message certainty is the degree to which the resulting

message obtained through a transaction describes the true state of the world. Message certainty comprises

of two factors, completeness and equivocality. Message completeness refers to conditions in which the

message is truncated. Message equivocality refers to conditions in which the message can have more than

one interpretation [12]. Message equivocality is most often caused by missing cues that would otherwise

indicate which interpretation is correct [12]. Figure 4.4 shows the relationship of completeness and

equivocality to message certainty. We will assume that the agent will resolve all message uncertainty

before the transaction completes. Message uncertainty is reduced through additional interaction such as

asking for clarification and the severity of the time penalty for message uncertainty is a function of both

agents’ abilities to resolve the message uncertainty over the given social medium. We will assume that any

additional interaction is part of the single transaction and that any additional interaction merely lengthens

the time to completion.

Figure 4.4a. Incompleteness in message certainty Figure 4.4b. Equivocality in message certainty

The choice of media affects the message certainty [12]. To better understand the relationship between

perceived cost and social media, we must look at the attributes of the social media that affect message

certainty, and hence, the agent’s cost estimation. The classification framework described below maps

social media along three semi-independent axes, each of which provide implications for cost perception.

Figure 4.5 shows several common social media mapped according to their synchronicity, directness, and

social presence.

The first axis, synchronicity, addresses whether social interaction occurs synchronously or asynchronously.

A synchronous system is real time where there is no delay in communication from either party. A face-to-

face meetings, phone conversations, and real-time discussions in chat-rooms are examples of synchronous

interaction. An asynchronous system involves delayed transmission of communicative messages. Email and

postings on newsgroups are asynchronous systems because the message may not be read immediately. In

general, asynchronous social systems have the potential for high associated time-costs due to the delay in

transmission and reception.

S A B C G

S A B

C

G

E

D

F ?

?

24

The second axis for evaluating social systems is directness. Svensson defines directness as the capacity for

mutual communication [46]. By this Svensson means the ability to reply – to reciprocate – to the initial

message. Most social interaction is direct because messages are exchanged in sequences that are related

contextually. Indirect interaction occurs when there is no mechanism for reciprocation such as when

information is obtained through a collaborative filtering system [17] or from a prerecorded message such as

a TV commercial. Directness tends to reduce the message equivocality and message uncertainty because

reciprocation can be used to refine and clarify the message. However, if the interaction is occurring through

an asynchronous system, the time-costs for reciprocation can build up. Indirect systems can have very low

associated time-costs because the interaction can be encapsulated as a single message where the knowledge

is completely and unequivocally transferred, requiring no reciprocation resulting in a short transaction. In

cases of indirect communication, the format of the message itself has a strong affect on message

uncertainty, as any ambiguity or incompleteness cannot be resolved without reciprocation.

The third axis is social presence. Social presence is degree of salience of another person in a social

interaction. Social presence is strengthened by feelings of immediacy and intimacy during interpersonal

communication [44] and is strongly correlated with effective learning [37] and persuasion [7]. Information

can be exchanged in parallel across any number of channels such as eye contact, gestures, and voice

inflection. Social systems that restrict the use of parallel channels tend to score low on social presence

MUDs, IRC

Newsgroups, Email

Collaborative filterers

CVEs, Real-word dialog

Real-world observations of others

TV Commercial

Direct

Indirect

Asynchronous Synchronous

Increasing Social Presence

Figure 4.5. Social media mapped onto the classification framework

25

scales. Not surprisingly, email, newsgroups, and MUDs, which restrict communication to a textual channel,

score low on social presence scales. Whether or not social presence is a desirable trait for a social system

depends on the nature of the communicative task. High social presence systems have been found to reduce

message equivocality during tasks involving high message equivocality [9]. Likewise, low social presence

systems have been found to reduce message incompleteness during tasks involving large amounts of

information transfer [9].

4.4. Limitations of the Computational Model

The primary limitation of the computational model is that it only considers the cognitive cost of planning.

Whether the agent chooses memory, the environment, or social interaction is based on an estimate of

cognitive effort, in terms of time, although the model could easily be extended to consider other measures

of cognitive effort. The model, as it currently stands, does not take into consideration physical resource

usage involved in navigation, such as the physical effort exerted or fuel spent achieving a state where social

media are available if social interaction is determined to be the best source of navigation knowledge. An

agent that faithfully follows the computational model of social navigation might decide that the best

strategy is to communicate when the distance to the nearest instance of a social medium is farther than the

distance to the goal itself. Such a strategy, while cognitively the most cost-effective, would be disastrously

expensive in terms of the physical effort of navigation. In order to correct for this limitation, physical costs

must be incorporated into the decision-making process of the agent. There are two possible approaches to

extending the computational model to account for physical effort: incorporation of physical cost into the

cognitive costs determined in the remote directories, or adding an additional decision-making stage. The

first alternative, incorporating physical cost into the remote directories, requires the directories in long-term

memory to be aware of physical factors such as the distance to reach an instance of a social medium and

how resources are expended during local and remote directory transactions. Unfortunately, it is not clear

how perceived physical costs, other than elapsed time, compare to the perceived cognitive costs already

considered by the directories. The second alternative, adding an additional decision-making stage to the

model, would allow the directories to behave as previously discussed, choosing the lowest-cost source of

knowledge based on perceived cognitive effort. The local processor (see figure 4.3) would then evaluate

the choice of knowledge source based on physical cost and expected resource usage. This approach

involves an additional amount of interaction between the local processor and the directories in local

memory that is not entirely consistent with the HIP model of decision-making.

26

5. PLANNING AND EXECUTION IN SOCIAL NAVIGATION

Interleaving planning and execution is a computational technique that captures the best of both classical

planning and reactive planning and can be of value to agents performing navigational tasks. While the

computational model of social navigation specifies where and how to acquire knowledge about the

environment and about the navigation task, it does not sufficiently address the issue of planning and

execution. The acquisition of knowledge through the knowledge-in-the-head, knowledge-in-others’-heads,

and the environment and the use of that knowledge can be thought of the planning aspects of navigation. In

any reasonably complex and uncertain environment, the agent will not be expected to acquire all the

information to build a full plan in only one location – execution must be performed when the agent needs to

acquire information that is not locally available.

5.1. Extending the Computational Model with Decision Theory

The social medium used to interact with others has been shown to have a direct impact on the perceived

cost of social communication, and hence, a direct impact on how browsing strategy is formulated. Since

the formulate browsing strategy stage of navigation is fundamentally a decision-making step, it can be

beneficial to discuss the choice of strategy in decision-theoretic terms. Casting strategy choice into

decision-theoretic terms will allow us to build computational software agents that can make rational

choices about how to navigate in social environments. Dean et al. [13] used decision-theoretic techniques

to design a control system for their explorative robot, HUEY. HUEY is a robotic agent that could build a

map representation of an uncertain environment and then use that representation to choose an appropriate

route through the environment, from start to goal. HUEY uses decision theory to choose between a path

through well-mapped portions of the environment and a potentially shorter path through a partially

unknown portion of the environment [13]. While HUEY does not consider sources of navigational

knowledge other than the map representation stored in memory or sensory interaction with the

environment, its use of decision theory to choose between navigation based on memory and navigation

based on reactivity is similar to using decision theory to make rational choices about planning and

execution in social environments.

Decision theory is used to make optimally rational decisions when an agent is faced with a set of two or

more actions to be chosen from. Each action is associated with a set of outcomes, which are states of the

world that could occur if a given action is selected. The agent assigns an expected utility value to each

outcome for each action and then chooses an action based on analysis of the expected utility values. The

expected utility for a given outcome is computed by assigning an arbitrary worth value to the outcome.

This value is an estimate of how valuable the agent believes being in the given state is. Furthermore, each

outcome has a certain probability of occurring if the corresponding action is chosen. The expected utility

value is the result of multiplying the worth value by the probability of the outcome occurring.

27

In the domain of navigation, an action is to navigate to a particular state in the world, X [13]. (Dean et al.

refer to these actions as point-to-point traversal tasks). The outcomes are the states of the world that the

agent might end up in when it tries to navigate to X. The desired outcome, of course, is X, but due to the

uncertainty of the world, some other state might be achieved instead. If the agent can assign expected

utility values to different states in the world, the agent can decide whether to plan or to execute any partial

plan by looking at the utility of planning and the utility of executing the partial plan. Each state of the

world can be thought of as having an associated cost. This is the cost of gathering information in that state.

Information can be gathered from the environment, from memory, or from others’ memories. Thus the cost

of being in a state is the cost of performing the cheapest transaction – the transaction that will be made if

the agent were in that state. Planning can be thought of as navigating to the current state – the agent has

chosen to use the resources available in the current state. Execution can be thought of as navigating to a

state described in the partial plan built by the agent. Navigation is only necessary when the plan is

complete or when the current state does not yield enough information for the agent to generate a complete

plan.

The perceived costs assigned to states of the world must first be converted into worth values. Costs are

considered to be inverses of worth values in decision theory, so conversion is straightforward. Expected

utilities can be computed by multiplying the estimated probability of reaching the state in question by the

worth value that we computed from cost. The probability of reaching the current state is, of course, one.

Performing this translation for desired outcomes and undesired outcomes of every action will result in a

table of expected utilities from which a rational choice of action can be performed. Figure 5.1 shows a

sample expected utility table.

Actions Desired Outcome Undesired Outcome

Navigation to current location (plan)

( )( )11 00 AA CEU = ( )( )01 11 AA CEU =

Navigation to B ( )( )BBB PCEU 00 1= ( )( )BBB PCEU −= 11 11

Navigation to C ( )( )CCC PCEU 00 1= ( )( )CCC PCEU −= 11 11

… … …

Figure 5.1. A sample expected utilities table

The computational model of social navigation is a reasonable way to describe the decision-making

processes behind navigating when social opportunities are available since the model is based on social

psychology. Decision-theory is a common technique for implementing computational decision-making

processes. Williamson and Hanks [51] describe a planning extension, PYRRHUS, which uses a decision-

theoretic approach to select a plan from the set of feasible plans that optimizes the use of resources usage.

In experiments using PYRRHUS, fuel, money, and time are all defined as resources in a modified

28

Truckworld [25] domain. Trucks must deliver shipments from one city to where fuel is consumed by the

journey and some roads are toll roads, requiring money to be spent. As in section 4.4, there are limitations

to the model that could be overcome by adding consideration for physical resources such as time, fuel, and

distance. Adding consideration for resource usage to the directories in local memory is one alternative for

addressing this limitation and would not affect the decision-theoretic extension to the computational model

of social navigation; the physical cost component would merely be added to the calculation of the expected

utility. Again, it is not clear how physical costs compare to cognitive costs, so some trial and error would

be needed to ensure that both aspects of perceived effort are scaled appropriately.

5.2. Extending the Computational Model with Congregations

An alternative to decision-theory is to use a representation used to describe the behaviors of distributed

multiagent systems, called congregations [5, 6]. A congregation is a group of agents that have chosen to

come together to a common place because other agents with complimentary services are more accessible

[5, 6]. In congregations multiagents must work together to solve problems where no single agent has all

the resources necessary. There is, therefore, a payoff involved being in a location where an agent can

readily access the resources of other agents. However, since some agents are better than others, the issue

becomes one of finding a congregation that maximizes ones own effectiveness [5, 6]. The cost of initially

seeking out other agents to congregate with will, in the long run, result in more cost-effective operations [5,

6]. Agents are free to join and remove themselves from congregations at any time.

Since the presence of other agents in the world and the ability to communicate with them in order to

acquire navigational knowledge had a direct affect on navigation, it is reasonable to model social

navigation as a collection of distributed agents. Social navigation can be viewed as an attempt to maximize

utility by briefly forming congregations with other agents that have the navigational resources required for

an agent to successfully complete the navigation task. Congregations are formed at points designated as

loci [6]. In the domain of social navigation, a locus can be considered any location state in the world where

an agent can communicate through social media with one or more other agents. The agent computes the

expected payoff over time related to being in different loci in the world, where payoff is directly related to

the utility of being in a congregation. However, due to the nature of the task, congregations must be short-

lived because the social navigation agent must eventually move in order to reach its goal. The problem of

planning and execution can thus be cast into the problem of deciding when to congregate and when to leave

a congregation.

The decision of when to congregate and when to leave a congregation is performed by looking at the long-

term payoff for each possibility and choosing the one with the highest long-term payoff. In this sense, the

congregation model is similar to decision-theory as discussed in section 5.1. Payoffs can be considered as

expected utility with a time component. Since a payoff value is contingent on other agents in the

congregation, the exact payoff value of a congregation cannot be known until the agent reaches the locus

29

[6]. Since the exact payoff value of a congregation cannot be known in advance, the agent must estimate

the expected payoff from what it knows about the loci in the world [6], which, in the domain of social

navigation, are defined by the social media that is available. This is consistent with the model of social

navigation in that a decision to interact with other agents is, in part, determined by the agent’s estimate of

the cognitive cost of interacting through a social medium.

When a social navigation agent is in a congregation – the agent is in a location where it has access, either

directly or indirectly, to other agents – it can, through illocution, receive additional knowledge resources,

adding to its plan. However, unlike the congregations discussed in [5] and [6], there are diminishing

returns; eventually the other agents in the congregation will no longer be able to help. As the payoff

expected from being in a congregation diminishes, an agent must choose when the payoff gained by finding

another congregation outweighs the payoff of the currently formed congregation. Diminishing utility

causes a destabilization of the environment as agents are pushed to move and form new congregations [5].

30

6. APPLICATIONS OF THE COMPUTATIONAL MODEL AND

CLASSIFICATION FRAMEWORK

The computational model of social navigation and its accompanying framework for classifying social

systems have two implications for use. First, the classification framework can guide the development and

deployment of new social media systems. Second, the model itself can be used to evaluate the design of

social environments by gauging the performance of simulated navigational software agents.

The success of social media will be determined by its positioning in the environment and its perceived cost

of usage. For example, the success of the telephone can be attributed to its ubiquitous nature and the ease

and efficiency. The ability to use the telephone to communicate with others quickly vastly offsets any

uncertainty attributed to the leanness of the medium. While the computational model of social navigation

says little about how to position instances of social media throughout the environment, it can help predict

whether the cost-effectiveness of using the medium will play a factor. If one understands how

synchronicity, directness, and social presence affect the perceived cost of using social media for a given

population, one can select the social media with the most desirable traits to place in the environment.

Similarly, if social media do not exist that already have the most desirable traits, new social systems can be

designed that have the most appropriate values of synchronicity, directness, and social presence. The

evaluation of social environments can indicate whether social navigation is sufficiently supported and, if

not, how the environment can be redesigned to better support user tasks and preferences.

6.1. The TRAILGUIDE System

The TRAILGUIDE system is a recommender system for navigation on the World Wide Web that was

designed using the classification framework for social media. The World Wide Web, due to its large,

unstructured nature can prove to be a challenging environment to navigate through [16]. Erickson [21] and

Dieberger [18] have both documented how social navigation can be beneficial to the Web and how users

have attempted to introduce social navigation tools to the Web through personal Web pages and

recommender systems. Upon classifying various social media on the Web such as personal Web pages,

recommender systems, newsgroup hyperlinks, and email hyperlinks in terms of synchronicity, directness,

and social presence we find a common trend. First and foremost, social media used for Web navigation are

all asynchronous and extremely low in social presence. The low social presence of these media is

attributed to the text-based nature of the communication channel. The inherent leanness of text-based

communication has been well documented [12, 9, 1]. Furthermore, the social navigation tools directly built

into the Web, such as personal Web pages and recommender systems tend to be indirect.

Asynchronicity and indirectness, by them selves, tend to increase the perceived cost of using a particular

communication medium. However, when a medium is both asynchronous and indirect, the perceived cost

of using that medium is dependent on the completeness and equivocality of the message being transferred.

31

If the message is complete and unequivocal – the directions are not missing any steps and those steps are

worded in an unambiguous fashion – there will be no need for reciprocation and therefore an asynchronous

and indirect message can be very quickly read and used. The low social presence in social media on the

Web was determined to be a large part of the cost of using social navigation on the Web. TRAILGUIDE

was designed to be an asynchronous, indirect recommender system with high social presence.

TRAILGUIDE is a system developed to maintain a repository of the recorded experiences on the Web for

oneself and for others. One user can recommend paths through the hyperlinks of the World Wide Web by

recording one’s own navigation and annotating certain aspects of what is viewed along the way. These

paths can be followed by other users and can learn from the paths. TRAILGUIDE consists of an authoring

tool and a play back tool. The authoring tool records the movements of the user as he browses from one

page to another; links that are selected for navigation are recorded as trail markers. Trail markers act as

signposts pointing the way that the user chose to go as well as an optional annotation of the reason the link

was chosen. When the user finds a Web page with interesting content, the content can be highlighted and

another trail marker will be created to store an annotation typed in by the user. In this manner,

TRAILGUIDE records a user’s experiences within and between pages on the Web. Advanced authoring

tools allow an author to write conditional statements into the trail markers. If the viewer wishes to know

more about a tangential topic that has been recorded by the author, he or she can follow any branching trail

and be able to jump back to the main trail at any time.

Digital information has no history [49]. For semantic spaces such as the World Wide Web, there are no

traces or cues to indicate a history of usage of online documents. Thus, navigation problem solving must

be approached by the problem-solver as if he were the first to make use of the online documents [49]. Most

recommender systems on the Web, such as collaborative filterers, are explicit in that they tell you exactly

where others with similar profiles have navigated. A class of recommender system, dubbed read wear,

attempts to mark digital documents with usage information. Read wear marks sections of digital documents

with wear-and-tear to indicate how often these portions of the document have been read [26]. This

technique can be applied to the World Wide Web by marking commonly accessed hyperlinks. The

FOOTPRINTS system [49] is a read wear recommender system that captures personal histories of Web

usage and compiles contextual visualizations of the semantic spaces. WALDEN’S PATHS [23, 42, 43]

also provides contextual guidance through the World Wide Web by allowing peers and teachers to author

paths that will take the user through a linear sequence of related documents. TRAILGUIDE is similar to

FOOTPRINTS and WALDEN’S PATHS in that it is a tool for recording experiences for the purpose of

assisting others with the navigation of the World Wide Web. TRAILGUIDE differs from other existing

recommender systems in that it attempts to use animated avatars to raise social presence in the same way

that CVEs increase social presence over textual MUDs.

Before anyone can begin recording with TRAILGUIDE, that person must first register a 2D-avatar body to

represent herself with. Preferably this avatar body should visually represent the user’s character,

32

personality, or role with respect to those who will view the users. When a user plays back a recorded trail,

instead of getting a static list of trail markers, he is presented with the author’s avatar, which proceeds to

gesture, emote, and speak the recorded thoughts of the author. Figure 6.1 shows TRAILGUIDE in use with

a playback in progress. During playback, the avatar indicates the links that the author followed but the

viewer does not need to select the links himself, the avatar can do it for him. Interaction with the avatar

mainly involves indicating that the avatar should continue the presentation (or go back to a previous point);

the avatar pauses after each trail marker is presented to allow the user to pursue her own thoughts or read

through Web page material at her own pace. Additional interaction takes place if conditionals and branches

are scripted into the trail, in which case the avatar can ask the user questions and receive limited feedback.

If the viewer chooses to, she can navigate away from the pre-recorded trail. The viewer can return to the

trail by asking the avatar to bring her back to the last played point or by navigating back on to the trail

manually. If the viewer navigates on to part of the trail not yet viewed (or any other trail in the repository),

TRAILGUIDE can pick up at that point or return the viewer to the last played point.

Figure 6.1. A Screenshot of TRAILGUIDE in use with playback in progress

33

A wide repertoire of communicative behaviors is available to be authored into the trail and acted out by the

avatar during playback. Possible behaviors include the more mundane pointing, moving, and glancing as

well as more emotionally motivated behaviors such as expressing pleasure, satisfaction, excitement, or

disappointment. With the use of avatars, TRAILGUIDE hopes to elevate the perception of social

presence above that of the traditional recommender system despite the fact that the true communicative

act is both asynchronous and indirect. Despite the asynchronous and indirect nature of communication

between author and reader through the TRAILGUIDE system, informal observations have indicated that

readers become affectively engaged with the presenting avatar. Such phenomena have been observed in

other domains [31].

6.2. A Testbed for Evaluating Navigability of Social Environments

The prediction of usage of a social medium relies on application of the classification framework for social

media. To use the computational model itself, we can evaluate entire social environments in terms of how

well they support social navigation. To evaluate a social environment for navigability, we needed a way

simulate an arbitrary navigation environment. The environment to be navigated with or without social

navigational aid could be semantic, such as a database or the World Wide Web, or spatial, such as a MUD,

CVE, or a real city. In order to simulate as wide a range of environments as possible, we created MUNE, a

Multi-User Navigation Environment based on the concept of the MUD. A MUD is a textual virtual

environment that users can navigate through and interact with other users and objects. MUD worlds

typically are designed around geography, which is divided into “rooms,” although these rooms can be

described in such a way as to appear as a road or an open field. Rooms are self-contained; there is no

interaction between events in one room and another adjacent room, including the ability to see into an

adjacent room.

MUNE is designed to be used by software agents instead of humans, so MUNE communicates according to

an expressive communication protocol easily parsed by computer agents but not human-readable. These

software agents can be any programs that understand the MUNE protocol and is able to make decisions

about how to navigate and interact with the virtual world described within. Human users can also access

the MUNE server through the use of a special client program that translates the computer-readable

protocols to natural language. MUNE is based on a MUD, but is designed as a testbed for navigation

problem solving. Emphasis is placed on descriptions of the world modeled within MUNE, but other

common features of MUDs, such as complex identity and dialog management, are excluded.

MUNE, unlike most other MUDs, does not assume a geographical world; its rooms can be designed to look

like nodes in a semantic environment. Instead of links between rooms being described as “east” or “north,”

the links can be described as hyperlinks or other such mechanisms for navigation. MUNE’s flexibility

stems from its simplicity and its extensibility. MUNE only knows about room descriptions and links

between rooms and user locations but allows for additional world descriptors. MUNE does attempt to

34

understand additional world descriptors, but merely passes this information along to any client programs

that might understand the descriptors. For example, in the real world, a person standing on top of a hill is

able to see for miles around. While it is possible for the room representing the top of the hill to be

described in such a way as to give the appearance of being able to see for miles around, a software agent

would need a more formal description. Such formal descriptions can be coded into the hilltop room as long

as the software agents understand the extension to the basic protocol. Figures 6.2a through 6.2c show

example worlds that can be created in MUNE. Figure 6.1a shows a conventional, spatial world based on

geographic terrain. A segment of code describing a particular room in this world is given. The “LINK”

descriptor tells MUNE which actions are legal in a given room. Figure 6.2b shows the same world with an

extension that allows software agents to see one or more rooms ahead. The gray circle designates what an

agent can see from the hilltop room. MUNE will ignore the “LOOKAHEAD” descriptor but send it to

clients that might be able to use that information for navigation. Figure 6.3c shows a semantic world based

on pages that might be found on the World Wide Web. Notice that the links are described as hyperlinks

instead of conventional spatial tags such as “east” or “north.”

Beyond the basic face-to-face interactions that can occur in a MUD, MUNE allows objects to be scripted

into the world that can facilitate communication such as phones, Email, etc. Unfortunately one limitation

of MUNE is that social interaction can only occur directly. In order to simulate indirect forms of

communication, such as recommender systems on the Web, non-navigational agents can be designed to

emulate the behavior of such a social navigation tool.

LOCATION ID Hilltop END-ID DESCRIPTION You are on a hilltop… END-DESCRIPTION LINK COMMAND south END-COMMAND DESTINATION Hilly Path END-DESTINATION END-LINK … END-LOCATION

Figure 6.2a. An example spatial MUNE world with description

Hilly Path

Hilltop D

B

G

A C

F E

Bottom of Path

R

S

T

H I J K L

M

N

O

P Q

35

LOCATION ID Hilltop END-ID DESCRIPTION You are on a hilltop… END-DESCRIPTION LINK COMMAND south END-COMMAND DESTINATION Hilly Path END-DESTINATION END-LINK … LOOKAHEAD Bottom of Path … END-LOOKAHEAD LOOKAHEAD Hilly Path … END-LOOKAHEAD LOOKAHEAD D … END-LOOKAHEAD LOOKAHEAD E … END-LOOKAHEAD … END-LOCATION

Figure 6.2b. An example spatial MUNE world with extensions to the description

LOCATION ID Home Page END-ID DESCRIPTION Welcome to my home page… END-DESCRIPTION LINK COMMAND link to NCSU END-COMMAND DESTINATION NCSU Home Page END-DESTINATION END-LINK … END-LOCATION

Figure 6.2c. An example semantic MUNE world with description

Inside the MUNE world, software agents can be given goals to navigate to. There is no limit to how the

software agents are designed to carry out their navigational tasks, including, but not limited to, depth-first

search, reactive search, or search using social navigation. MUNE is flexible enough that a large variety of

environments can be simulated. Social navigation agents can be tweaked to behave according to certain

preferences and preconceived notions about social interaction that real users of the simulated environment

might have. The navigability, both in terms of general navigation strategies as well as social navigation

strategies, of the environment can be evaluated by measuring and comparing the performance of social

navigation software agents, conventional navigation software agents, and human users in the simulated

MUD environment. Human users can interact with the simulated MUD environment through special client

software that converts MUNE world specification format into human readable room descriptions, although

human users may find the textual nature of MUDs more cumbersome to interact with than a visual spatial

Hilly Path

Hilltop D

B

G

A C

F E

Bottom of Path

R

S

T

H I J K L

M

N

O

P Q

Home Page

NCSU Home Page

Friend’s Home Page

Hobby Page

36

environment. Because the leanness of textual descriptions is limiting to human users and not to software

agents, it may not be possible to directly compare the performances of human users and software agents

operating within MUNE, unless the conditions are well controlled for or more expressive client programs

are developed. Once performance measures have been collected, the navigation environment can be easily

adjusted and social media can easily be re-distributed until a desirable level of performance is reached.

This approach to evaluation can be cumbersome because the navigation environment must be encoded in

MUNE world description format and software agents must be developed. A more general approach to

simulating a navigation environment is used to evaluate the computational model of social navigation in

section 7. In the generic approach, an environment is modeled by a set of parameters that express the

environments navigability without resorting to a room-level description of the world, although the results

expected from such a non-specific simulation will be much less detailed than the results that MUNE can

provide.

37

7. SIMULATING NAVIGATION IN SOCIAL ENVIRONMENTS

In order to evaluate the computational model of social navigation, we designed and developed a simulation

that will reveal how social awareness affects navigation. The simulation is based on the principal of

interleaved planning and execution described by the computational model and should provide a reasonable

approximation of agent behavior when confronted by a variety of environments and given various

preferences for social media usage. By analyzing the simulation results, we will be able to determine

whether the computational model is capable of generating reasonable behaviors in navigational agents. The

objective of this study is not to evaluate agents that search for optimal paths through the environment;

rather to evaluate how well an agent performs, compared to the baseline optimal path.

The simulation does not model specific agents or specific environments, but instead operates on classes of

environments and agents, defined by the parameterization of each. The simplicity of the design allows us

to analyze the broad patterns of behavior in a large variety of environments without getting bogged down

by the details of implementing and testing real agents and the design and construction of real environments.

A more complete study can be performed by developing software navigation agents for MUNE (see section

6.2). The richness of the MUNE environment can reveal more details about the specific behaviors and

routes chosen by the agents, allowing for a more detailed analysis of agent behavior in social environments.

This level of detailed analysis is beyond the scope of the work presented here.

The simulation is intended to provide insight into behavioral patterns of agents that use the computational

model of social navigation proposed above. Through the simulation, we can vary the environment through

which a simulated agent would navigate and aspects of the simulated agent itself to determine how

different combinations of parameters will affect overall performance. We expect to see patterns of

behavior that are plausibly explained by the computational model and by anecdotal evidence in the social

navigation literature. Erickson [21] and Dieberger [16] both refer to situations in which Web users will

seek out tools for social navigation. Other studies of strategies for navigation [24, 27] show subjects

performing navigation that is not directly goal-related in order to ease the cognitive demands of an

unknown environment. It would be of interest to determine what features of the environment and of the

social media used in social navigation have effects on agent performance when navigating through

unfamiliar spaces. In specific, under what conditions are social navigational strategies appropriate and

when should more traditional navigation strategies be adopted?

7.1. Simulation Design

The simulation is a way of quantifying the ways in which social interaction affects navigation. The

simulation engine assumes an arbitrary environment in which navigation can be represented as following

links through an undirected graph. At each node in the graph, one or more out-link is optimal in that it is

part of the path that will take the agent from the current node to the goal in the shortest number of steps.

38

There can be several out-links that are part of sub-optimal, but valid, paths to the goal. There are also out-

links, referred to as “dead ends” that take the agent to nodes that are not part of any path to the goal unless

backtracking is performed. The complexity and uncertainty of the environment can be parameterized by

these six variables without knowing the degree of the graph. For each of these three classes of out-links,

there is an associated cost of following an out-link. In a perfectly ambiguous environment, for example a

maze, every out-link has the same probability of being chosen. If a maze is comprised of four-way

intersections, then the probability of choosing the optimal out-link is 1/3 (we do not consider the link from

which we arrived at the current node). In a less ambiguous environment, this probability increases. Given

the probabilities of choosing a class of out-links at any given node in the graph and the optimal length from

start state to goal state, we can compute the expected number of nodes visited using the binomial statistical

distribution. This expected number of nodes visited gives us a baseline for which to compare results using

social navigation. The navigation environment is thus parameterized by the following six values: optimal

distance, Dopt, the probability of choosing an optimal out-link, Poptlink, the probability of choosing a dead-

end, Pdeadlink (the probability of choosing a sub-optimal link is implicit since all three probabilities must sum

to one), the cost of following an optimal out-link, Coptlink, the cost of following a sub-optimal link, Csublink,

and the cost of following a dead-end link, Cdeadlink.

Socialization during navigation is parameterized by four variables: the distribution of social media, Dsocial,

the perceived cost of using social media, Csocial, the actual time it takes to use the social medium, Tsocial, and

the number of steps generated through illocution, L. Social media can be other people, phones, Email, or

any communication technology through which illocution can occur. On a university campus, the

distribution of other people through the environment might be quite high. In a city environment,

telephones are distributed such that one can be found every few miles. Alternatively, Email might have a

very low distribution because publicly accessible computer terminals are quite rare. Distribution of social

media is measured by the average number of nodes one must visit before encountering another instance of a

social medium.

Following the computational model of social navigation, the simulated agent makes at each stage whether

to use social navigation or whether to reactively search for the goal. If reactive search is chosen, the

simulation engine computes the expected number of nodes visited until the agent must make another

decision. While the model calls for a decision to be made at every node, in practicality the decision will

not vary until something in the environment has changed significantly, which is based on the distribution of

social media in the environment. If social navigation is chosen, the simulation engine computes the

expected number of nodes visited before reaching an instance of the social medium. This is extraneous

navigation that is not directly goal-related. The number of steps generated through illocution is then

subtracted off the total distance to the goal as goal-directed navigation; we assume that once directions are

received, no error in navigation is made while the plan is being executed. The cycle of planning and

execution is repeated iteratively until the goal is reached.

39

The simulation engine is simpler than the computational model in that it only allows for one remote

directory at a time. Experimentation with social navigation agents operating in MUNE environments has

shown that the agent will always transact with the perceived cheapest cost knowledge source, all other

things being equal. In a more complex domain, different remote memories will be non-overlapping in the

portions of the navigation environment they are knowledgeable about. For the purposes of the simulation

we assume that transactions with all remote memories cost the same and that there is always a remote

memory that can assist the navigational agent as long as some instance of a social medium is locally

available.

The simulation engine, as a result, returns five measured values: the distance the agent is expected to travel,

Dfinal, the elapsed time the agent is expected to arrive at the goal, Tfinal, the percentage of the baseline

(without social navigation) distance traveled when social navigation is used, RD, the percentage of the

baseline elapsed time when social navigation is used, RT, and the number of times the agent chooses to

interact socially, Nsocial. Expected distance and expected time traveled are two commonly measured

variables in navigation research from which we can deduce the level of difficulty navigation in the

parameterized environment will pose the agent. The percentage measurements, RD and RT, tell us whether

social navigation is more efficient in certain circumstances, how much more efficient social navigation is,

and, as other parameters change, how much slower the distance and time traveled using social navigation

changes with respect to distance and time traveled without social navigation. Nsocial is redundant because it

grows proportionally with Dfinal, but turns out to be very important when analyzing certain anomalies

observed in the other measured values.

7.2. Simulation Procedure

Of the ten parameters to the simulation engine, two were chosen at a time and varied across a range of

reasonable values while all other parameters are held constant. The data set generated by this pairing can

be examined for any interactions between the two parameters. There are 45 possible M-by-N analyses that

can be made, although not all combinations prove to be useful. Of the M-by-N analyses that were chosen,

each one is run several times with different sets of constants in order to assess whether there are any

additional interactions Different sets of constants are used and the data set is regenerated in order to assess

whether there are any 3-way interactions. Data sets were graphed in various ways and observed for

interesting patterns. Initially, the simulation engine was run without any of the social navigation

parameters. This enabled us to ensure that there were no unexpected patterns that arose from various

combinations of environmental parameterizations as well as provided baseline patterns that we could

contrast to patterns involving social navigation.

40

7.3. Simulation Results

Running the simulation engine without the four social navigation parameters provides predictable results.

As the probability of choosing the optimal out-link decreases, the expected distance and time to reach the

goal increases proportionally. The probability of choosing a sub-optimal out-link and the probability of

choosing a dead-end link determines which of the respective cost parameters dominates the growth. The

relationship is shown in figure 7.1a. Establishing this, we can begin to look at the effects that social

interaction has on navigation. Social navigation did decrease the expected distance the agent must travel to

reach the goal, but only in certain circumstances. When conditions were right, the use of social interaction

increases the expected travel distance, as indicated by RD > 1. The remainder of this section is broken into

a discussion of the parameters that cause monotonic increases in the four dependent variables and the

discussion of the parameters that cause non-monotonic increases in the dependent variables.

Figure 7.1a. Contour graph of distance traveled affected by the probability of choosing an

optimal out-link and the probability of choosing a dead-end out-link

Figure 7.1b. Relationship between the time ratio and the number of social interactions

7.3.1. Monotonic observations

The most interesting observation is that none of the parameters cause a strict monotonic increase in Dfinal,

RD, or RT. Instead, all graphs of data sets have interesting jags and dips (see figures 7.2a, 7.3, and 7.4).

Upon closer examination of the data, it becomes apparent that Nsocial, which increases in a step pattern as

the number of decision points increases, causes these anomalous features (see figure 7.1b); Nsocial can only

assume integer values. In most circumstances, an increase in the number of social interactions will cause a

dramatic decrease in Dfinal, RD, Tfinal, and RT. The largest jag at the top of figure 7.2a is the result of Nsocial

dropping to zero because large values of L reduce the number of decision points – the first illocution

41

provides the agent with a large enough plan that illocution need not be performed again. Once we

compensate for these anomalies, we begin to see monotonic patterns. The monotonic patterns are

enumerated below.

• The distribution of social media throughout the environment, Dsocial, has a profound effect on the

distance the agent will travel, Dfinal. As the distance the agent must travel to congregate increases, the

overall distance to travel to reach the goal will increase. In fact, Dsocial does not need to grow very large at

all before social navigation is less efficient than blind search. The only tempering factor is the uncertainty

of the environment. When Poptlink is close to 1.0 – the uncertainty of the environment is very low – the

distribution has a reduced impact on Dfinal because fewer errors are made while the agent is trying to

congregate. Unfortunately, as the environment becomes less uncertain, the usefulness of congregation is

reduced, unless the usefulness of congregation, L, increases proportionally. See figures 7.2a and 7.2b.

Figure 7.2a. Contour graph of distance traveled affected by the social media

distribution and environmental uncertainty

Figure 7.2b. Contour of the distance ratio affected by the social media distribution and

environmental uncertainty

When uncertainty in the environment is high, low distribution of social media throughout the environment

can result in the agent following a very lengthy path. This occurs because the agent seeks out congregation

but expends a lot of energy doing so because the uncertainty of the environment impedes congregation. If

the uncertainty is lower, then the agent can congregate more easily and will end up traveling a shorter path.

However, as uncertainty decreases, the agent does not need to rely on social navigation as much; in an

environment with very low uncertainty, any navigation to congregate is extraneous because reactive search

can reveal the path to the goal just as easily or even more easily if the distribution of social media is low.

• The perceived cost of social interaction, Csocial, has very little effect on distance unless Csocial is very

large. When the cost of social interaction is very large, then the number of social interactions, Nsocial, will

quickly drop to zero. It appears that perceived cost is primarily a factor for choosing between the best

42

source of interaction and best social medium through which to conduct the interaction. In this simulation,

which only looks at the possibility of one social source at a time (assuming the given source is the best),

perceived cost has little effect. For the same reason that a high perceived cost drives the number of social

interactions down, cost does have an impact on the severity of the afore mentioned jags. The higher cost,

the more radical the jags can affect Dsocial and RD. This phenomenon is due to the fact that, as the agent

approaches the goal and reactivity becomes more cost-effective, a high perceived cost would push the agent

towards reactivity sooner causing greater disparity when Nsocial steps down.

• The actual time to use an instance of the social medium, Tsocial, predictably causes an increase in the

overall time to reach the goal, Tfinal. The actual time required to complete illocution is not factored into the

utility of planning because the agent cannot know this value. The only defense the agent has against high

and unwieldy actual transaction times is to estimate the time required and incorporate that into its perceived

cost of the social medium. Ideally, we would like to see perceived cost have some effect on Tfinal as Tsocial is

varied. In fact, we do find that, in general, if actual time is low, the agent benefits from assigning a low

cost to social media usage. Conversely, if actual time is very high, the agent benefits from assigning a high

cost to social media usage. Only when the actual time and cost are distinctly different do we see poor

performance from the simulated social navigation agent. Figure 7.3 shows the relationship actual

transaction time and perceived cost has on RD.

Figure 7.3. Contour graph of the time ratio affected by perceived cost and actual transaction time

• The number of plan steps acquired from illocution, L, has a direct inverse effect on distance traveled,

Dfinal and RD. Specifically, as the number of steps acquired decreases, the distance traveled increases

exponentially. This exponential curve is interrupted by fluctuations caused by the sudden step-downs of

Nsocial, as mentioned before. An interesting point is that when Nsocial = 1, the increase of Dfinal is constant as

L decreases, causing an interesting discontinuity in figure 7.4. When Nsocial > 1, as Nsocial periodically steps

down, the exponential growth of Dfinal and RD becomes less pronounced until the Nsocial = 1 step where

43

growth becomes linear. There is another discontinuity in the graph in figure 7.4 where perceived cost of

using the social medium is very high. At this point, no social interactions are being performed so RD is

constant at 1.

Nsocial = 1

Nsocial > 1

Nsocial = 0

Figure 7.4. Contour graph of the distance ratio affected by plan steps and perceived cost

When the length of the plan acquired through illocution is small, the distance that the agent travels can be

very large. This is because the distance that the agent travels to reach a congregation is greater than the

distance it can travel based on the plan steps it receives from the congregation. Only when plan length is

very small do we see very large distances traveled. The distance traveled tapers off exponentially as plan

length increases towards a more useful size. An anomaly occurs when, through various reasons, the agent

only congregates once. When this situation occurs, plan length is linearly proportional to the distance the

agent travels. The acquired plan will take the agent a given distance and the remainder of the path will be

determined reactively. Of course, when plan length is high, less reactivity is needed and the distance the

agent ends up traveling will be lower. When the agent never seeks out congregation (because the perceived

cost of social interaction is too high) the entire path is determined reactively.

7.3.2. Non-monotonic observations

The non-monotonic observations are more interesting because they indicate complex interactions between

two or more parameters that would not be immediately obvious without generating large data sets from the

simulation. The most interesting interaction occurs between L, Dsocial, and Csocial. The first thing that is

noticed is that when Dsocial > L, RD > 1. This observation makes sense because if Dsocial is large and L is

small, then the agent is going to expend more energy getting assistance than it gets back from obtaining

assistance. The non-monotonous relationship is evident; as L increases the distribution of social media has

a decreasing effect on Dfinal and RD. This lessening of effect occurs more rapidly than linear and, at some

point, actually reverses itself. Choosing any value of Dsocial on figure 7.5a (figures 7.5a through 7.5c have

44

been smoothed for illustrative purposes) and moving horizontally from left to right, we see RD decrease

non-linearly and then begin to increase. This non-linear decrease and reversal is non-intuitive since one

would guess that a large L could only be a benefit to a navigation agent.

Figure 7.5a. Contour graph of the distance ratio affected by plan length and social media

distribution when perceived cost is low

Figure 7.5b. Contour graph of the distance ratio affected by plan length and social media

distribution when perceived cost is medial

Figure 7.5c. Contour graph of the distance ratio affected by plan length and social media distribution when perceived cost is high

However, upon closer inspection, we can deduce that, when L is large but not larger than Dopt, the agent

will be delivered very close to the goal and, being so close to the goal, determine that reactive search is

more cost effective than social navigation. A tendency to favor reactive search when near the goal is the

cause of the non-linear decrease, and may actually be the best strategy. The reversal occurs because when

Dsocial is small enough that social navigation is still the favored strategy even when very close to the goal.

Social navigation, at this point, just adds more extraneous distance to Dfinal than is gained because the agent

is so close to the goal. The perceived cost of social interaction, Csocial, plays key part in this interaction,

because the higher the cost of social interaction is perceived to be by the agent, the less likely it is to use

45

social navigation at the very end of the task, mitigating the effects of the rise. As Csocial approaches Dopt,

the contours straighten out, as seen in figures 7.5a through 7.5c.

When the distance the agent must travel to use an instance of a social medium is greater than the number of

steps acquired through illocution, social navigation is less efficient than reactive search; the agent expends

more energy reaching the congregation than it receives as reward for congregating. When social navigation

is beneficial, the perceived cost of social interaction and the plan length acquired through illocution have an

interesting relationship. Increasing the plan length has a definite advantage up until a certain point when

increasing the plan length starts to reduce the effectiveness of social navigation. When plan length is high,

the agent will be delivered very close to the goal. But if the perceived cost of congregating is low, the

agent will seek out additional social interaction even when it is in clear sight of the goal and reactivity will

actually serve the agent better. Increasing the perceived cost of social interaction will mitigate this effect

because the agent will be less likely to choose to congregate when the goal is nearby. Of course, increasing

the perceived cost of social interaction too high will cause the agent to never congregate.

These results are consistent with independent observations that human navigators, when faced with

navigation in an unfamiliar environment, will adjust their goals to seek out advance information through

social interaction, even when such goals took them off of the direct path [24]. Human navigators prefer

social navigation to map reading because social instructions narrow the space and provide contextually

relevant information that cannot be acquired easily without prior experience [27]. Erickson [21] also points

out the common phenomenon of searching for someone who would know how to reach the goal instead of

searching for the goal itself in both real-world navigation and navigation on the World Wide Web. Social

navigation has been cited as preferred over blind search and over non-social navigation aids such as search

engines because of their foreign and daunting natures [21]. One important finding from the simulation of

social navigation is that there are circumstances when performance using social navigation does not strictly

dominate performance using asocial navigation. The question remains whether using social navigation

under inappropriate circumstances is avoidable given a more expressive model of the perceived costs of

social interaction. The current model of social navigation is limited by its disregard of physical resource

consumption and consideration of such physical costs by the computational model of social navigation may

required. The data obtained from the simulation suggest that when an agent’s perception of social

interaction costs reasonably reflect the nature of the environment and of social media, navigation can be

made efficient by choosing the navigation strategy, social or asocial, that best enhances performance.

46

8. CONCLUSIONS

The simulation of navigation in social environments has shown to produce reasonable results. The

distribution of social media through the environment impacts the distance the agent is expected to travel to

reach the goal. The distance traveled to reach an instance of a social medium is assumed to be distance

traveled that is not goal-oriented and therefore extraneous. In a highly uncertain environment, reaching an

instance of a social medium can be as problematic as progressing towards the goal. When the distribution

is very low and the agent must travel a long way to reach the social medium, distances traveled can

increase radically. This increase in expected distance due to the distribution of social media in the

environment is offset by the usefulness of social interaction, as measured by the number of plan steps

acquired by the agent through illocution. These plan steps are considered to be accurate and reduce the

uncertainty of the portion of the environment they describe. The simulation shows that the greater the

number of steps acquired through illocution the more efficient the navigation is, as indicated by the

percentage of the distance expected without social navigation.

One would assume from the commonality of social navigation in human behavior [21] that social

navigation is strictly more efficient than asocial navigational practices, such as reactive search. The

simulation, however, has revealed conditions in which social navigation is inferior to reactive search. One

such situation occurs when the number of plan steps acquired through illocution is not greater than the

distance the agent must travel to perform the illocution. This is a reasonable condition for social navigation

to produce poor results; social navigation assumes helpfulness and is bound to fail when this assumption is

not met. Another condition in which social navigation proves to be less efficient than reactive navigation is

when the number of steps acquired through illocution is high, but not high enough to deliver the agent all

the way to the goal. While counter-intuitive, we see from the simulation that the agent, when very close to

the goal, the strategy – social or asocial – used for navigation is critical to the final efficiency. The

simulation suggests that the agent should rely on reactive navigation when close to the goal because the

usefulness of social interaction is diminished, unless the distribution of social media is extremely high

(requiring on a short distance to reach an instance of the social medium). We view this as an indication that

the limitations discussed in section 4.4 do exist and can be effectively handled by extending the model, in

some fashion, to account for resources used by the agent.

We have also seen that an agent that can reasonably estimate the cost of using social media will perform

much better than an agent that cannot. The estimate does not have to be accurate for performance to be

good, only that the agent estimates the cost to be high when the time to use the social medium is in fact

high or that the agent estimates the cost to be low when the time to use the social medium is in fact low. It

is fortunate that the agent only requires a reasonable estimate and not an accurate estimate because human

use of heuristics to make decisions relies on approximate situation assessment.

47

Finally we have seen that social navigation is more beneficial in environments that are highly uncertain

than in environments that are not very uncertain at all. Within highly uncertain environments, benefit from

social navigation comes only when instances of social media are widely available. In an uncertain

environment, reactive search will cause the agent to make many mistakes that will be costly in terms of the

distance the agent travels while backtracking. Social navigation can reduce the number of mistakes the

agent makes, but only if the agent can reach an instance of a social medium without too much difficulty. If

it is too difficult for the agent to reach an instance of a social medium, the agent will not gain any

advantage and may actually be penalized for choosing social interaction.

While the simulation has shown that the computational model of social navigation can produce plausible

results, it remains to be evaluated in more detail. A more detailed analysis of social navigation is not

within the scope of this research, but the framework for a more detailed experiment is already in place. The

MUNE system can be used to extract more detailed behavior patterns from agents that use social

navigation. Due to the nature of MUNE, these agents can be software agents, implementing the

computational model of social navigation described above, or they can be human agents. We expect any

detailed analysis to fall within the patterns observed through our more limited simulation. Through

resolution of the limitations of the computational model, variations may occur. The limitations to the

computational model, as noted at the time of writing, involve the assignment of cost and utility. Simulation

results have hinted at this limitation and the possible correction by including perceived physical costs, such

as distance to be traveled, to the perceived transaction costs in the computational model. Further work is

needed in order to determine if the cost framework for choosing between planning and execution can

sufficiently integrate all the factors that an agent might consider when choosing between planning – using

the environment, memory, and others’ experiences – and execution.

48

9. LIST OF REFERENCES

1. Adrianson, L. & Hjelmquist, E. (1999). Group processes in solving two problems: face-to-face and

computer mediated communication. Behaviour and Information Technology, 18(3), 179-198.

2. Ambrose-Ingerson, J.A. & Steel, S. (1988). Integrating planning, execution, and monitoring.

Proceedings of the Seventh National Conference on Artificial Intelligence, Saint Paul, MN, 1988, pp.

83-88.

3. Arkin, R. (1989). Navigational path planning for a vision-based mobile robot. Robotica, 7, 49-63.

4. Arkin, R. (1989). Towards the unification of navigational planning and reactive control.

Proceedings of the AAAI Spring Symposium on Robot Navigation, Stanford University, March 1989.

5. Brooks, C.H. & Durfee, E.H. (1999). Congregation formation in information economies. AAAI

Workshop on Aritificial Intelligence in Electronic Commerce, July 1999.

6. Brooks, C.H., Durfee, E.H., & Armstrong, A. (2000). An introduction to congregating in multiagent

systems. Proceedings of the Fourth International Conference on Multiagent Systems, Boston MA,

July 2000.

7. Burgoon, J.K., Birk, T., & Pfau, M. (1990). Nonverbal behaviors, persuasion and credibility. Human

Communication Research, 17(1), 140-169.

8. Bush, V. (July 1945). As we may think. The Atlantic Monthly, 176(1), 101-108.

9. Chidambaram, L. & Jones, B. (1993). Impact of communication and computer support on group

perceptions and performance: a comparison of face-to-face and dispersed meetings. MIS Quarterly,

17(4), 465-488.

10. Cohen, P.R. & Levesque, H.J. (1990). Rational interation as the basis for communication. In P.R.

Cohen, K. Morgan, & M.E. Pollack (eds.) Intentions in Communication. Cambridge, MA: The MIT

Press. pp. 221-255.

11. Dahlbäck, N. (1998). On spaces and navigation in and out of the computer. In N. Dahlbäck (Ed.)

Exploring Navigation; Towards a Framework for Design and Evaulation of Navigation in Electronic

Spaces. Technical Report T98:01. Kista, Sweden: Swedish Institute of Computer Science, pp. 15-

29.

12. Daft, R.L. & Lengel, R.H. (1986). Organizational information requirements, media richness, and

structural design. Management Science, 32(5), 554-571.

13. Dean, T., Basye, K., Chekaluk, R., Hyun, S., Lejter, M., & Randazza, M. (1990). Coping with

uncertainty in a control system for navigation and exploration. Proceedings of the Eighth National

Conference on Artificial Intelligence, Boston MA, July 1990.

49

14. Dean, T., Kaelbling, L., Kirman, J., & Nicholson, A. (1995). Planning under time constraints in

stochastic domains. Artificial Intelligence, 76(1-2). 35-74.

15. Dieberger, A. (1996). Browsing the WWW by Interacting with a Textual Virtual Environment -- A

Framework for Experimenting with Navigational Metaphors. Proceedings of HYPERTEXT '96,

Bethesda MD, 1996. 170-179.

16. Dieberger, A. (1997). Supporting social navigation on the World Wide Web. International Journal

of Human-Computer Studies, 46(6), 815-825.

17. Dieberger, A. (1998). Social connotations of spatial metaphors and their influence on (direct) social

navigation. In K. Höök, A. Munro, & D. Benyon (Eds.) Workshop on Personalized and Social

Navigation in Information Space. Technical Report T98:02, Swedish Institute of Computer Science,

Kista, Sweden, 1998.

18. Dieberger, A. (1999). Social connotations of space in the design for virtual communities and social

navigation. In A.J. Munro, K. Höök, & D. Benyon (Eds.), Social Navigation of Information Space

(pp. 15-32). London: Springer-Verlag.

19. Dourish, P., (1999). Where the footprints lead: tracking down other roles for social navigation. In

A.J. Munro, K. Höök, & D. Benyon (Eds.), Social Navigation of Information Space (pp. 15-32).

London: Springer-Verlag.

20. Dourish, P. & Chalmers, M. (1994). Running out of space: models of information navigation.

Proceedings of HCI’94, Glasgow, August 1994.

21. Erickson, T. (1996). The World-Wide-Web as social hypertext. Communications of the ACM, 39(1),

15-17.

22. Etzioni, O., Hanks, S., Weld, D.S., Draper, D., Lesh, N., & Williamson, M. (1992). An approach to

planning with incomplete information. Proceedings of the International Conference on Principals of

Knowledge Representation and Reasoning, 1992, pp. 115-125.

23. Furuta, R., Shipman, F.M. III, Marshall, C.C., Brenner, D., & Hsieh, H. (1997). Hypertext Paths and

the World-Wide Web: Experiences with Walden's Paths. Proceedings of HYPERTEXT '97,

Southhampton UK, 1997. 167-176.

24. Gordon, D.A. & Wood, H.C. (1970). How drivers locate unfamiliar addresses – an experiment in

route finding, Public Roads, 36(2), 44-47.

25. Hanks, S., Pollack, M., Cohen, P. (1993). Benchmarks, testbeds, controlled experimentation, and the

design of agent architectures. AI Magazine, 13(4).

26. Hill, W.C., Hollan, J.D., Wroblewski, D., & McCandless, T. (1992). Edit wear and read wear.

Proceedings of the Conference on Human Factors in Computing Systems, Monterey, CA, 1992, pp.

3-9.

50

27. Höök, K. (1998). Route guidance issues; verbal versus map instructions, and route choices. In N.

Dahlbäck (Ed.) Exploring Navigation; Towards a Framework for Design and Evaulation of

Navigation in Electronic Spaces. Technical Report T98:01, Swedish Institute of Computer Science,

Kista, Sweden, 1998.

28. Hutchins, E. (1995). Cognition in the Wild. Cambridge, MA: MIT Press.

29. Jul, S., & Furnas, G. (1997). Navigation in electronic worlds. SIGCHI Bulletin, 29(4).

30. Kuipers, B. (1982). The “map in the head” metaphor. Environment and Behavior, 14(2), 202-220.

31. Lester, J.C., Converse, S., Kahler, S., Barlow, T., Stone, B., & Bhogal, R. (1997). The Persona

Effect: affective impact of animated pedagogical agents. In Proceedings of CHI '97 (Atlanta GA,

1997), ACM Press, 359-366.

32. Moreland, R.L. (1999). Transactive memory: learning who knows what in work groups and

organizations. In L.L. Thompson, J.M. Levine, & D.M. Messick (Eds.) Shared Cognition in

Organizations. Mahwah, NJ: Lawrence Erlbaum Associates. pp. 3-31.

33. Newell, A. & Simon, H.A. (1972). Human Problem Solving. Englewood Cliffs, NJ: Prentice Hall,

inc.

34. Nilsson, N.J. (1984). Shakey the robot. Technical Report 323, AI Center, SRI International, Menlo

Park, CA.

35. Norman, D.A. (1990). The Design of Everyday Things. New York: Doubleday.

36. Resnick, P. & Varian, H.R. (1997). Recommender systems. Communications of the ACM, 40(3),

56-58.

37. Richmond, V.P., Gorham, J.S., & McCroskey, J.C. (1987). The relationship between selected

immediacy behaviors and cognitive learning. In M. McLaughlin (ed.) Communication Yearbook, 10.

Beverly Hills: Sage, pp. 574-590.

38. Riedl, M.O. (2001). A computational model and classification framework for social navigation.

Proceedings of the International Conference on Intelligent User Interfaces, Santa Fe, NM, January

2001.

39. Roberts, J.M. (1964). The self-management of cultures. In W.H. Goodenough (Ed.) Explorations in

Cultural Anthropology: Essays in Honor of George Peter Murdock. New York: McGraw-Hill. pp.

433-454.

40. Rosen, D.L. & Olshavsky, R.W. (1987). The dual role of informational social influence:

implications for marketing management. Journal of Business Research, 15(2), 123-144.

41. Schoppers, M.J. (1989). Universal plans for reactive robots in unpredictable domains. Proceedings

of the Tenth International Joint Conference on Artificial Intelligence, Menlo Park, CA., 1987.

51

42. Shipman, F.M III, Marchall, C.C., Furuta, R., Brenner, D.A., Hsieh, H.W., & Kumar, V. (1996).

Creating Educational Guided Paths over the World-Wide-Web. Educational Telecommunications,

1996: Proceedings of ED-TELECOM '96, pp. 326-331.

43. Shipman, F.M. III, Furuta, R., Brenner, D., Chung, C., & Hsieh, H. (1998). Using Paths in the

Classroom: Experiences and Adaptations. Proceedings of the Ninth ACM Conference on

HYPERTEXT '98, Pittsburgh PA, 1998. 267-270.

44. Short, J., Williams, E., & Christie, B. (1976). The Social Psychology of Telecommunications.

London: John Wiley & Sons.

45. Spence, R. (1999). A framework for navigation. International Journal of Human-Computer Studies,

51, 919-945.

46. Svensson, M. (1998). Social navigation. In N. Dahlbäck (Ed.) Exploring Navigation; Towards a

Framework for Design and Evaulation of Navigation in Electronic Spaces. Technical Report

T98:01. Kista, Sweden: Swedish Institute of Computer Science, pp. 73-88.

47. Walsh, J.P. & Ungston, G.R. (1991). Organizational memory. Academy of Management Review,

16(1), 57-91.

48. Wegner, D.M. (1995). A Computer Network Model of Human Transactive Memory. Social

Cognition, 13(3), 319-339.

49. Wexelblat, A. & Maes, P. (1999). Footprints: history-rich tools for information foraging.

Proceedings of the CHI 99 Conference on Human Factors in Computing Systems, Pittsburgh PA,

1999. 270-277.

50. Wickens, C.D. & Hollands, J.G. (2000). Engineering Psychology and Human Performance, 3rd

Edition. Upper Saddle River, NJ: Prentice Hall.

51. Williamson, M. & Hanks, S. (1994). Optimal planning with a goal-directed utility model.

Proceedings of the Second International Conference on Artificial Intelligence Planning Systems

(AIPS-94), June 1994, pp. 176-181.