Upload
others
View
1
Download
0
Embed Size (px)
Citation preview
CAP6671 Intelligent Systems
Lecture 2:Multiagent Systems
Instructor: Dr. Gita SukthankarEmail: [email protected]
Schedule: T & Th 9:00-10:15amLocation: HEC 302
Office Hours (in HEC 232):T & Th 10:30am-12
2CAP6671: Dr. Gita Sukthankar
Homework
Reading: D. Nau, Current Trends in Automated Planning, AI Magazine (posted on web site once I scan it in)
3CAP6671: Dr. Gita Sukthankar
What is an agent?
4CAP6671: Dr. Gita Sukthankar
What is an agent?
MAS: “loosely coupled network of problem solvers that interact to solve problems that are beyond the individual capabilities or knowledge of each problem solver”Belief-Desire-Intention model of agency:
SituatedGoal-directedReactiveSocial
Framework for constructing BDI agents: Procedural Reasoning System
5CAP6671: Dr. Gita Sukthankar
Deliberative vs. Reactive
What is the difference?
6CAP6671: Dr. Gita Sukthankar
Deliberative vs. Reactive
Deliberative:Attempt to attain a set of goalsThis paper only really talks about deliberative agents.
Each individual agent is an intelligent system.
Reactive: Respond to the current environment using a stimulus/response paradigmAgents are individually dumb and “intelligence” is an emergent property of the way that the system is connected.
7CAP6671: Dr. Gita Sukthankar
RETSINA Agent
KQML: knowledge query machine language
8CAP6671: Dr. Gita Sukthankar
Multi-Agent Organization
User 1 User 2 User u
InfoSource 1Info
Source 1
Interface Agent 1Interface Agent 1 Interface Agent 2Interface Agent 2 Interface Agent iInterface Agent i
Task Agent 1Task Agent 1 Task Agent 2Task Agent 2 Task Agent tTask Agent t
MiddleAgent 2MiddleAgent 2
Info Agent nInfo Agent n
InfoSource 2Info
Source 2
InfoSource mInfo
Source m
Goal and TaskSpecifications Results
SolutionsTasks
Info & ServiceRequests
Information IntegrationConflict Resolution Replies
Advertisements
Info Agent 1Info Agent 1
Queries
Answers
distributed adaptive collections of agents that help users by retrieving information, providing advice and anticipate user's information needs.
9CAP6671: Dr. Gita Sukthankar
Matchmaker
MatchmakerRequester
Provider 1 Provider n
Request for service
Contact information of providers that match the
request Advertisementof capabilities
Delegation of service
Results of service request
Yellow-page system
10CAP6671: Dr. Gita Sukthankar
MAS Coordination
Teamwork: Agents share a common set of goals and each contributes to the fulfillment of these goals through teamwork. No explicit modeling of individual agent utility
Coalitions: Agents seek to maximize individual utility and group utility (coalition stability is an issue)
Coordination: Agents pursue their individual goals and utilities; coordinationwith others is done only to avoid harmful interactions (e.g. traffic)
Negotiation: Agents seek to maximize their individual utility but are willingto compromise (i.e. better off if they reach agreement than not)
Game Theoretic Interactions: Agents seek to maximize individual utility while taking into consideration other’s options
Adversarial Interactions/Zero Sum Games: Agents seek to maximize own utility while minimizing utility of opponent
11CAP6671: Dr. Gita Sukthankar
Characteristics of an MAS
Incomplete informationNo global controlData is decentralizedComputation is asynchronous
12CAP6671: Dr. Gita Sukthankar
Why use an MAS?
13CAP6671: Dr. Gita Sukthankar
Why use an MAS?
Solve problems too huge for one agentInteroperation of existing legacy systemsPreserve privacy of user dataUse spatially separated information sourcesDistributed expertise
14CAP6671: Dr. Gita Sukthankar
Aircraft-serviceAir-Force base 3
Hangar computer
Aircraft Maintenance
Aircraft-serviceAir-Force base 2
Hangar computer
Formagent
Aircraft-serviceAir-Force base 1
Hangar computer
Historyagent
Historyagent
Historyagent
Formagent
Form202A
202DB
202DB
202DB
Manuals ManualsManuals
Manualsagent
Manualsagent
Manualsagent
Requestrelevant historicalrepairs
Requestrelevant manualpages
Engineer A’s workstation
Engineer B’s workstation
Requestrelevant historicalrepairs
Form202B
Form202A
Sent bymechanic
Sent bymechanic
Form202B
Sent to mechanic
To mechanic
15CAP6671: Dr. Gita Sukthankar
Aircraft Maintenance
Shehory, Sukthankar, and Sycara 1998Why use an MAS?
User interface agents needed to be lightweight and run on either a tablet computer or desktopUtilize legacy systems that the military already had in place
16CAP6671: Dr. Gita Sukthankar
Agent-based Programming
Does it improve system performance?
17CAP6671: Dr. Gita Sukthankar
Claims
Computational efficiencyReliabilityExtensibilityRobustnessMaintainabilityResponsivenessFlexibilityReuseDoes agent-based programming really offer all those advantages?
18CAP6671: Dr. Gita Sukthankar
Examples of MAS Applications
Sensor netsSmart buildingsShopping bots/financial agentsInformation-gatheringManufacturingPersonal information management agents
Calendar scheduling
Air traffic controlVirtual humans/embodied conversational agents
19CAP6671: Dr. Gita Sukthankar
Information Fusion with MAS
Information Information AgentAgent
Middle Middle AgentAgent
Logistics Logistics (Task) (Task) AgentAgent
Interface Interface AgentAgent
Fusion Fusion (Task) (Task) AgentAgent
Information Information AgentAgent
Information Information AgentAgent
Information Information AgentAgent
Information Information AgentAgent
Information Information AgentAgent
Information Information AgentAgent
Information Information AgentAgent
Information Information AgentAgent
Information Information AgentAgent
Information Information AgentAgent
Middle Middle AgentAgent
Middle Middle AgentAgent
Middle Middle AgentAgent
Middle Middle AgentAgent
Logistics Logistics (Task) (Task) AgentAgent
Interface Interface AgentAgent
Fusion Fusion (Task) (Task) AgentAgent
Customized Views for
Different Roles
Fused Data from different
Battlefield regions
Reconnaissance Data
Multilevel fused intelligence reports
20CAP6671: Dr. Gita Sukthankar
Research Issues
Task allocationCNP, auction-based frameworks
Multiagent planning: resolving sub-gal interactions
Decentralized MDP, PGP (partial global planning)Teamwork
Modeling other agentsPlan recognition
Resource/Communication limitationsDistributed constraint optimization
Adaptation and learningMulti-agent RL
21CAP6671: Dr. Gita Sukthankar
What remains unsolved?
When this article was written:Lack of systematic methodologyIndustrial-strength MAS toolkits
Now:Larger-scale systemsHuman-agent interfacesGlobal interoperabilityDynamic planning/problem-solvingLifelong learning
22CAP6671: Dr. Gita Sukthankar
Vision: 10 Years Out
•Memory augmentation •Personalized services•Nano-robots/smart matter•Self-updating media (content and context).•Ambient anytime anywhere information access (safe mobile info environments for people)•Highly realistic agents that inhabit large virtual environments