Autonomous Agents Overview
Dec 21, 2015
Topics
• Theories: logic based formalisms for the explanation, analysis, or specification of autonomous agents.
• Languages: agent-based programming languages.
• Architectures: integration of different components into a coherent control framework for an individual agent.
Topics
• Multi-agent architectures: methodologies and architectures for group of agents (could be from different architectures)
• Agent modeling: modeling other agents’ behavior or mental state from the perspective of an individual agent
• Agent capabilities
• Agent testbeds and evaluation
What is an agent?
• Weak:– Autonomy– Social ability– Reactivity– Pro-activities
• Strong:– Mental properties such as knowledge, belief,
intention, obligation– Emotional
• Others attributes: mobility, veracity, benevolence, rationality
Agent Theories
• How to conceptualize agents?
• What properties should agents have?
• How to formally represent and reason about agent properties?
Agent Theories
• Definition: an agent theory is a specification for an agent. Formalisms for representing and reasoning
about agent properties
• Starting point: agent = entity ‘which appears to be the subject of beliefs, desires, etc.’
Intentional system
• An intentional system whose behavior can be predicted by the method of attributing belief, desires, and rational acumen
• Proved that can be used to describe almost everything
• Good as an abstract tool for describing, explaining, and predicting the behavior of complex systems
Intentional system - Examples
• One studies hard because one wants to get good GPA.
• One takes the course ‘cs579-robotic’ because one believes that it will be fun.
• One takes the course ‘cs579-robotic’ because there is no 500-level course offered.
• One takes the course ‘cs579-robotic’ because one believes that the course is easy
Agent Attitudes
• Information attitudes: related to the information that an agent has about the environment – Belief– Knowledge
• Pro-attitudes: guide the agent’s actions– Desire– Intention– Obligation– Commitment– Choice
• An agent should be represented in terms of at least one info-attitude and one pro-attitude. Why?
Representing intentional notions
Representing Jan believes Cronos is the father of Zeusnaïve translation into FOL: Believe(Jan, Father(Zeus,Cronos)) Problems:
1. No nested predicate 2. Zeus = Jupiter Believe(Jan, Father(Jupiter,Cronos)) [Wrong]
Conclusion: FOL is not suitable since intention is context dependent.
Possible World Semantics
• Hintikka: 1962 – Agent’s belief can be characterized as a set of possible worlds.
• Example: – A door opener robot: door is closed, lock needs to be
unlocked but the robot does not know if the lock is unlocked or not – two possibilities:
• {closed, locked} • {closed, unlocked}
– Card player (poker): ?– UNIX Ping command: ?
Possible World Semantics
• Each world represents a state that the agent believes it might be in given what it knows.
• Each world is called a epistemic alternative.• The agent believes in something is true in all
possible worlds.• Problem: logical omniscience – agent believes
all the logical consequences of its belief impossible to compute.
Alternatives to PWS
• Levesque – belief and awareness: explicit belief (small) from implicit belief (large).– No nested belief– The notion of a situation is unclear – Under certain situation: unrealistic prediction
• Konolige – the deduction model: modeling the belief of a symbolic AI system (database of beliefs and an inference system).– Simple
Others
• Meta-language: one in which it is possible to represent the properties of another language – Problem: inconsistency
• Pro-attitudes: goals and desires – adapting possible world semantics to model goals and desires – Problem: side effects
Theory of agency
• Realistic agent: – combination of different components– dynamic aspect
• Moore – knowledge and action: study the problem of knowledge precondition for actions– I needs to know the telephone number of my friend
Enrico in order to call him.– I can find the telephone number in the telephone
book.– I needs to know that the course is easy before I sign
up for it
Theory of agency
• Cohen and Levesque – belief and goal: originally developed as a pre-requisite for a theory of speech acts but proved very useful in analysis of conflict and cooperation in multi-agent diaglogue, cooperative problem solving
Theory of agency
• Rao and Georgeff – belief, desire, intention (BDI) architecture: logical framework for agent theory based on BDI, used a branching model of time
• Singh: logics for representing intention, belief, knowledge, know-how, communication in a branching-time framework
Theory of agency
• Werner: general model of agency based on work in economics, game theory, situated automate, situated semantics, philosophy.
• Wooldridge: modeling multi-agent system
Agent Architectures
• Construction of computer systems with properties specified by an agent theory.
• Three well-know architectures:– Deliberative– Reactive– Hybrid
Deliberative architecture
• View agent as a particular type of knowledge based system – known as symbolic AI
• Contains an explicit represented, symbolic model of the world
• Decision is made via logical reasoning (pattern matching, symbolic manipulation)
• Properties:– Attractive from the logical point of view– High computational complexity (FOL: not decidable,
with modalities: highly undecidable)
Sense• Assimilate Sensing results
Reasoning• Symbolic representation of the world• Determine what to do next
Act• Execute the action generatedby the reasoningmodule
ENVIRONMENT
Deliberative architecture in picture
Deliberative architecture
• Examples:– Planning agents: a planner is an essential component
of any artificial agent• Main problem: intractability – addressed by techniques such
as hierarchical, non-linear planning.
– IRMA (Intelligent Resource-bounded machine architecture): explicit representations of BDI & planning library, a reasoner, opportunity analyser, a filtering process, a deliberation process (mainly: reduced the time to deliberate)
Deliberative architecture
• HOMER: a prototype of an agent with linguistic capability, planning and acting capability.
• GRATE*: layered architecture in which the behavior of an agent is guided by the mental attitudes of beliefs, desires, intentions, and joint intention.
Reactive architecture
• Proposed to overcome the weakness of symbolic AI
• Main features: – does not include any kind of central symbolic
world model– does not use complex reasoning
Sense• Assimilate Sensing results
Reasoning• Determine whatto do next
Act• Execute the action generatedby the reasoningmodule
ENVIRONMENT
Reactive architecture in picture
Reactive architecture
• Brook - behavior language: subsumption architecture– Hierarchy of task-accomplishing behaviors– Each behavior competes with others – Lower layer represents more primitive task
and has precedence over upper layers– Very simple– Demonstrate that it can do a lot – Multiple subsumption agents
Reactive architecture
• Arge and Chapman – PENGI: most everyday activity is ‘routine’ – Once learned, a task becomes routine and
can be executed with little or no modificationRoutines can be compiled into a program and
then updated from time to time (e.g. after new tasks are added)
Reactive architecture
• Rosenschein and Kaelbling - Situated automata– Agent is specified in declarative terms which
are then compiled into digital machine – Correctness of the machine can be proved– No symbol manipulation in situated automata,
thus efficient
• Maes – Agent network architecture: an agent is a network of competency modules
Hybrid architecture
• Combine deliberative and reactive architecture – exploit the best out of the two
• Georgeff and Lansky – Procedural Reasoning System: BDI & plan library, explicit symbolic representation of BDI– Beliefs are facts – FOL– Desires are represented by behavior– Each plan in the plan library is associated with
invocation condition reactive – Intention – the set of currently active plans
Active
PlanLibrary
P1: Invocation I1
Pn: Invocation In
Belief: FOL
Desire: System beha.
Intention:
Pi: Invocation Ii
Pj: Invocation Ij
SystemInterpreter
Environment
PRS in picture
Hybrid architecture
• Ferguson – TOURINGMACHINES: – Perception and action subsystem – interact directly
with the environment– Control framework system: three control layers – each
is independent, activity producing, concurrently executing process
• Reactive layer (response to events that happen too quickly for other to response)
• Planning layer (select plan, actions to achieve goal)• Modeling layer (symbolic representation, use to resolve goal
conflict)
Hybrid architecture
• Burmeister et al. – COSY: hybrid BDI with features of PRS and IRMA, for a multi-agent testbed called DASEDIS
• Mueller et at. – INTERRAP: layered architecture, each layer is divided into knowledge and control vertical part
Agent language
• A system that allows one to program hardware and software computer systems in terms of some of the concepts developed by agent theorists.
• Shoham – agent-oriented programming:– A logical system for defining the mental state of
agents– An interpreted programming language for
programming agents– An ‘agentification’ process, for compiling agent
program into low-level executable systems Agent0: first two features