AutoPilot Year 1 Results Principal Investigators: Jerry Stach E.K. Park University of Missouri - Kansas City
Dec 30, 2015
AutoPilot Year 1 Results
Principal Investigators:
Jerry Stach
E.K. Park
University of Missouri - Kansas City
Questions Posed By Sponsor
• 1. Decentralized Scaleable Trader (a) How do you maintain global information about a set of
available services without a central point of failure? (b) What are the Query and Update costs of a decentralized
Trader?
University of Missouri - Kansas City
Questions Posed By Sponsor
• 2. Agent Health Monitor
Investigate ways to monitor large numbers of mobile/distributed agents with minimal effect on overall systems performance.
University of Missouri - Kansas City
Questions Posed By Sponsor
• 3. Mobile Agent Patterns As work on the first two categories progresses,
document any design patterns that emerge.
University of Missouri - Kansas City
Engineering Problems - Current System
• Only about 103 agents can run concurrently in the current network. The answers to the Sponsor’s questions must scale at least one order of magnitude:
• 104 concurrent mobile agents
• 104 service nodes with 10 – 20 service instances per node
• the number of Trader Places approaching 103
• 2*103 entries per directory.
University of Missouri - Kansas City
Underlying Causes
• The current limitation of 103 Agents implies Band Limiting• Band Limiting can Occur in the Service Place CPU if mean
processing time is high relative to agent arrivals• Band Limiting can occur in the Network as a function of
message intensity– focused overloads (agent Trader Place)– agent collaboration or management
University of Missouri - Kansas City
Band Limiting in the Service Place CPU
• Possible Strategies– Accelerate CPUs – Increase number of CPUs• Acceleration may not place the power in the right
locations• Increasing the number of Service Places increases
band width demand of the network
University of Missouri - Kansas City
Implications
• If arbitrary increases in the number of Service Places are to be avoided, concurrency is implied to maximize CPU utilization– agents should have capability to function as
autonomous distributed processes– mobility must include agent reasoning about local
congestion and distance
University of Missouri - Kansas City
Band Limiting in the Network
• Strategy– Minimize number of messages in the Agent Colony
• messages associated with accessing the Trader Place regarding Service offerings [Sponsor Question 1]
• messages associated with managing the colony of agents [Sponsor Question 2]
University of Missouri - Kansas City
Implication
• Minimizing the number of messages to the Trader Places implies some intermediate process in the Architecture that can parse the Trader Place vector for multiple agents at a Service Place. The agent must then be able to interpret the vector relative to its own preferences. The mobility decision is implied at the agent (lowest Architecture level) not at the Trader (highest Architecture level)
University of Missouri - Kansas City
Implication
• Minimizing the number of messages associated with population management implies being able to anticipate the location of a given agent to eliminate exhaustive search or suspension of the population. This implies higher levels of the Architecture must understand the mobility decision in order to locate agents in the network
University of Missouri - Kansas City
Problem Synthesis
• A multi - level Architecture is implied
Agent
Service Planner
Trader
University of Missouri - Kansas City
Problem Synthesis
• Migration is fundamental to the answers to Questions 1, 2– agents are situated and must move to a subsequent
Service Place based upon• service attributes of time, cost, quality (perception)
• a context of total moves that satisfies their service sequences (local optimization)
University of Missouri - Kansas City
Problem Synthesis
• Agent migrations should not adversely effect the network system– moves should be sensitive to network dynamics such
as local congestion and path length (global optimization)
– in the large, individual agent migration decisions should cause load leveling at Service Places and traffic distribution without a central authority (emergent behavior)
University of Missouri - Kansas City
Proposed AutoPilot Architecture
Agent
Service Planner
Trader
Topologist
Service Place
Trader Place
AutoPilot Network Router
University of Missouri - Kansas City
Research Implications:Distributed Artificial Intelligence
- there is an Artificial Intelligence Sub Problem
- there is a Distributed Processing sub problem
- there is a Network sub problem
University of Missouri - Kansas City
Artificial Intelligence Sub Problem
• Given a set of agent preferences for time to service, cost of service and quality of service, select the most desirable location from a set of possible locations that conform to the agent’s preferences
This is a multi-attribute programming problem
University of Missouri - Kansas City
Distributed Processing Sub Problem
• Given a set of Service Places and the service set of each Service Place, find an optimal assignment of Services to the Service Places subject to the Service Place environments
This is a multi-processor task assignment problem
University of Missouri - Kansas City
Network Sub Problem
• Given a sequence of Services specified by the agent’s work flow signature and a set of feasible Service Places, construct a optimal itinerary that minimizes total trip time
This is a graph theory problem (trip planning)
University of Missouri - Kansas City
Strategy
• Solve academic problems in a manner that produces engineering solutions as well as new knowledge
• Select solution techniques that integrate the three classes of sub problems
University of Missouri - Kansas City
Research Approach: Composition not Decomposition
• 1. Obtain a solution for the assignment of Services to Service Places
• 2. Obtain a solution to for agent's attribute based perception of Service Places
• 3. Integrate the results from (1) and (2) forming a mobility heuristic
• 4. Validate the heuristics by simulating situated multi-agents in a network of Service Places.
• 5. Formulate Trader Place Inquiry/Update Costs
University of Missouri - Kansas City
Overview of Assignment of Services to Service Places
• Each agent carries a work flow signature for the possible processing sequences of its task graph
AB
C
D
E
Work Flow Signature = A;B;(C+D);E
University of Missouri - Kansas City
Assignment of Services continued
• Build an interior graph of the signature. Weight interior edges with the payload size from taskI to taskJ
AB
C
D
E
1.0 2.83.6
.8
1.1
Input is initial agent payload and a scaling matrix
University of Missouri - Kansas City
Assignment of Services continued
• Connect each interior Service node to every Service Place supporting that service subject to the agent’s preference criteria
• Weight the edges from the Services to the Service Places by agent preference for the Service Place
University of Missouri - Kansas City
Assignment of Services continued
_serviceQuality_of Move. Possible
qAgent.w Software
cAgent.w Software
qAgent.w
rviceCost_of_se Move. Possible
qAgent.w Software
cAgent.w Software
cAgent.w
,
∗+
+∗
+
=
Software
Software
qix
University of Missouri - Kansas City
Assignment of Services continued
)(12
11
,,1
,, qiq
qrriqi x
nn
xn
w ? −
−−−
=
Weighting of edges from Services to Service Places
University of Missouri - Kansas City
Assignment of Services continued
A
B D
CE
C a,b
C b,d
C b,c
C d,e
C c,e
SP1SP2
SP3
Wa,sp1
Wa,sp3
Wb,sp1 Wd,sp2
We,sp3Wc,sp3
University of Missouri - Kansas City
Assignment of Services continued
A
B D
CE
C a,b
C b,d
C b,c
C d,e
C c,e
SP1
SP2
SP3
Wa,sp1
Wa,sp3
Wb,sp1
Wd,sp2
We,sp3
Wc,sp3
Not SP1
Find a minimum cut to the network - Services A,B,C are assigned to SP1
University of Missouri - Kansas City
D
E
C d,eSP2
SP3
Wd,sp2
We,sp3
Re-compute weights, find a new minimum cut, D is assigned to SP2, E is assigned to SP3
Assignment of Services continued
Final Service Assignments
SP1:= A,B,C
SP2 := D
SP3 := E
University of Missouri - Kansas City
Improving Performance
• We do not want to consider 10K Service Places for each agent– Observations
• several locations may be equivalent by agent perception of time to service, cost of service and quality of service
• if we could pick the Service Places to consider in the right order, we should assign all services in a relatively few iterations
University of Missouri - Kansas City
Improving Performance continued
• Leads to the multi attribute programming problem
• An agent perceives each Service Place by its attributes (time, cost,quality)
• If the agent could rank the Service Places by these attributes, we could generate equivalence classes of Service Places
University of Missouri - Kansas City
Improving Performance continued
Sp3SP5SP4 SP2SP1
SP5SP3SP2SP1SP4
SP3SP2SP1SP4SP5
Equivalence By Time Equivalence By Cost Equivalence By Quality
SP3 is an non-dominated Service Place in intersectionof the first equivalence class for each attribute.Have the graph algorithm consider SP3 first.
University of Missouri - Kansas City
Multi-Attribute Programming Problem
• We cannot use a linear weighting scheme to rank nodes because time, cost and quality do not normalize
• an agent’s constant perception of its environment is time
• the Topologist can provide the Service Planner the current geodasic to a Service Place (router interface)
University of Missouri - Kansas City
Multi-Attribute Programming continued
• Humans distort time by attributes – long car ride for a bargain is viewed as acceptable to
some limit of time– a one hour poor presentation is long– a two hour great movie is short• Why not let the agent distort time by the attributes of
Service Places?– Need an objective function
University of Missouri - Kansas City
Functions of Cost and Quality on Time
))(1()(
))(1()(
qc
qtq
qc
ctc
ww
wwqtf
www
wctf
+−=
+−=
University of Missouri - Kansas City
Quantifying agent perception
[ ]n_distortiomax
tftfserviceinitiatetotime
perception
qc
≤
+↔
=
)()(___
In AutoPilot we limit max_distortion to twice the diameter of the network so an agent perceives the timeto initiate service from nearly zero to twice the networkdiameter depending on its perception of the Service Place.
University of Missouri - Kansas City
Demonstrations of Research Results
• Description of Base Cases presented
• Results viewed by visual front end
• single agent simulations– link speeds are negligible
• heuristic search for service - equal preferences for time, cost, quality
• migration by preference for time
• migration by preference for cost
University of Missouri - Kansas City
Demonstrations of Research Results continued
• Multi-agent simulation– colony of 100 agents– all services offered on all nodes– arrival rates to network are high relative to processing
time– transmission times are negligible– hope to see second order network effects as emergent
behavior
University of Missouri - Kansas City
Network Second Order Effects as a result of multi-agent interaction
Emergent Colony Behavior• Under network loading individual agent decisions aggregate to Service Place load-leveling in the absence of any central network or Trader authority.
Expected Behavior Desired Behavior
Legend CPU UtilizationService Place Queue LengthService Place Agent Age
Accomplishments• 1. Obtain a solution for the assignment of Services to Service
Places Complete
• 2. Obtain a solution to for agent's attribute based perception of Service Places Complete
• 3. Integrate the results from (1) and (2) forming a mobility heuristic Complete
• 4. Validate the heuristics by simulating situated multi-agents in a network of Service Places.
– Partially Complete, base cases only, not fully debugged
• 5. Formulate Trader Place Inquiry/Update Costs– Equations presented in year end report
Proposed Research ActivitiesYear 2
• Focus on remaining Sponsor questions:
– 1(a). Decentralized Scaleable Trader
How do you maintain global information about a set of available services without a central point of failure?
–2. Agent Health Monitor Investigate ways to monitor large numbers of mobile/distributed
agents with minimal effect on overall systems performance subject to the Trader cost formula from Year 1 results.
University of Missouri - Kansas City
Proposed Research Activities Year 2 continued
• generalize the multi-attribute function for n attributes• fully debug the simulator and extend to accommodate a
definition of the Sponsor’s network (links/ number nodes)
• extend the simulator to include Trader Place update policies
(interval, random..)
University of Missouri - Kansas City
Proposed Research Activities Year 2 continued
• study the relationship between emergent behavior and the Trader Place update policy
• improve visualization post processor
• in first half year produce two journal papers on agent mobility– multi attribute programming solution– general formulation of agent mobility
University of Missouri - Kansas City
Proposed Research Activities Year 2 continued
• study the applicability of Artificial Life principles to agent mobility in large colonies.
University of Missouri - Kansas City