Distributed AI an overview
D Goforth - COSC 4117, fall Why distributed AI? situated expert – the importance of general knowledge and incorporation of distinct points of view – CYC human problem-solving teams with different expertise (and representations!) complexity of problems requires decomposition – OOP distributed problems – decentralized problem-solving – internet, air-traffic control
D Goforth - COSC 4117, fall Multi-agent systems parallel action at some level emergent structure chemical – pressure and temperature biological – bee hives mathematical – fractals artificial organization decentralized multi-agent systems emergent solution to problems
D Goforth - COSC 4117, fall Multi-agent systems agents in environment agents each interact with environment (perception, action) agents interact with each other levels of interaction vary independent influence through environment direct communication
D Goforth - COSC 4117, fall Multi-agent system problems agents have distinct / common goals independent competitive (can interfere with each other) cooperative (can help each other) collaborative agents have common goal one shot problems or ongoing survival
D Goforth - COSC 4117, fall Distributed systems – problem space amount of interaction between agents degree of commonality or conflict of goals single or ongoing operation
D Goforth - COSC 4117, fall Emergent solutions - examples efficient traffic flow based on actions of individual agents powerful search engine based on web-crawling agents just-in-time delivery and minimal inventory eBay
D Goforth - COSC 4117, fall Internet artificial environments distributed solutions – web crawlers artificial environments to enable distributed solutions – auction and bid software
D Goforth - COSC 4117, fall Internet artificial environments policy and common goals rules of environment agents act to achieve individual goals within rules achieve common policy goals also
D Goforth - COSC 4117, fall eBay environment parallel auctions – auction search engine extended but fixed bidding interval large potential bidding audience agents bidding agent
D Goforth - COSC 4117, fall Example – low cost telephone service in artificial market place current problem competition based on service plans hard to understand and compare constrains complexity of cost/service structure waste of resources on advertising (instead of cost reduction or service improvement) difficult for new service providers to enter market
D Goforth - COSC 4117, fall Low cost telephone service in automated negotiating environment two classes of agent: service providers customers telephones environment - phonecall marketplace intelligent telephone requests service service providers submit offers telephone selects one offer and connects to service provider market handles accounting and billing
D Goforth - COSC 4117, fall Low cost telephone service in automated negotiating environment advantages competitive on service and rate no service plans to understand since no long term commitment easy for service providers to change pricing easy for service providers to enter market intelligent telephone agent maximizes self interest (min cost for reqd service) service providers maximize self interest (maximize profit)
D Goforth - COSC 4117, fall Low cost telephone service in automated negotiating environment designing the environment how is bidding managed? goal get companies to bid the lowest price they can offer get companies NOT to bid strategically (bid maximum they think will win)
D Goforth - COSC 4117, fall Low cost telephone service in automated negotiating environment strategic bidding consider what others will bid operate customer agents to elicit offers from other service providers bid just less than competition how to suppress strategic bidding Vickreys mechanism lowest bid wins lowest bidder is paid at second lowest rate
D Goforth - COSC 4117, fall Vickreys mechanism example A bids to provide service at 10¢ / min B bids to provide service at 12¢ / min all other bids higher A wins contract, paid 12¢ / min rationale – incentive to relate bid to true cost no incentive to underbid (might win and have to provide service at a loss) no incentive to overbid (might lose unnecessarily and no gain in profit otherwise)
D Goforth - COSC 4117, fall Low cost telephone service in problem space no interaction between agents pure conflict between goals ongoing operation
D Goforth - COSC 4117, fall Example environments Electric power grids Robots on assembly line Bank transactions Traffic flow Distributed computing positions in problem space?
D Goforth - COSC 4117, fall What is DAI? AI (intelligent agent) game theory (interaction of agents) distributed computing
D Goforth - COSC 4117, fall Negotiation problem environment: communication between agents language of communication – protocols agents: goals tactics – using protocols to achieve goals how to achieve the best deal concessions, lies, threats
D Goforth - COSC 4117, fall Negotiation problem example domains Task-oriented domains State-oriented domains Worth-oriented domains
D Goforth - COSC 4117, fall Task-oriented domains Agents can act independently Agents cant interfere with each other Only incentive is possible cost reduction by cooperation (e.g., school boards sharing school bus routes)
D Goforth - COSC 4117, fall State-oriented domains Each agent has goal of environment in certain state Agents can interfere with each other – goal states in conflict or with mutual goal at high cost (limited resources) Incentive to negotiate – concede some goals; pay extra cost
D Goforth - COSC 4117, fall Worth-oriented domains generalized S-ODs – value function defines value of every state for agent possibility of efficient solutions with compromise – search model
D Goforth - COSC 4117, fall Negotiation problem example domains amount of interaction between agents degree of commonality or conflict of goals single or ongoing operation TOD SOD/WOD
D Goforth - COSC 4117, fall Negotiation mechanisms the negotiation system provided by the environment desirable properties of negotiation global optimality – policy goal efficiency – dont waste agent resources stability – no incentive to leave a deal distributed – no central authority required fairness – no preference based on external properties (not symmetry)