Information EconomiesSlide 2Open Markets of Independent AgentsEconomic MechanismsResearch AgendaChallenges of the WebResearch ApproachesA News Filtering EconomyModelModel (cont’d)Slide 11Slide 12Analysis of the SystemSystem with one brokerTwo-broker SystemSystem Behavior: Two-broker casePrice warsCan it be overcome?Large SystemLarge System (cont’d)Emergent BehaviorConclusions.1Information EconomiesBy J. Kephart, J.Hansen, D. Levine, B. Grosof, J. Sairamesh, R. Segal, “Emergent Behavior in Information Economies”.2Information EconomiesResearch AgendaInformation Filtering EconomyMonopolyDuopolyLarge SystemEmergent behavior in information economies.3Open Markets of Independent AgentsInternet evolves towards a milieu of Agents.Open, free-market information economy of agents engaged in economic activities.Inter-agent economic transactions.Is information economy inevitable?Increasingly important role of agents in e-commerce;Evolution of agents from transaction facilitators to decision-makers;Increasing autonomy and responsibility.4Economic MechanismsTraditionally developed for human agents.May not be directly applicable to agent communities:Agents make decisions and act at a much greater speed;Agents are much simpler then humans with respect to their governing rules;Less flexible;Lack in “common sense”.Do these differences matter?5Research Agenda“Is an information economy inherently capable of governing billions of software agents, and if so what are the minimal requirements on the infrastructure of such an economy and on the agents that populate it? ”Dynamical systems: nonlinear systems are capable of chaotic behavior.SI agents can be susceptible to unpredictable collective behavior.Efficient resource allocation as an emergent feature of a system.6Challenges of the WebOpen system.No global purpose.No global cooperation.No universal medium of exchange.No universal ontology.No universal set of agent types or algorithms.Will these features emerge?7Research ApproachesAnalyticalGame theory, mechanism design, etc.Multi-Agent Simulation 8A News Filtering EconomyAgents engaged in trading news articles.Simple behavioral rules.Focus on dynamical and economical aspects of the system;Assume that non-economic issues (security, transaction processing, etc.) have been solved.9ModelSource agent: publishes the articles.C consumer agents: want to buy articles.B broker agents: buy selected articles from the source and resell them to consumers.System infrastructure that provides computation and communication infrastructure for the agents.All the agents maximize their utility function.10Model (cont’d)SystemSource, {j}, PsBroker 1, {1j}, P1Broker 2, {2j}, P2Consumer 1, {1j}, VConsumer 3, {3j}, VConsumer 2, {2j}, VPTPT, PCPCPSPS11Model (cont’d)The source agent publishes an article at each time step t and assigns it to category j with probability j.The article is offered for sale to all brokers at PS. For each article sold to the broker the source agent pays PT transportation fee to the system.A broker decides whether or not to buy the article;The decision may be uncorrelated with the classification.For each evaluation the broker pays PC computation fee to the system.The broker’s decision method is parametrized through (random process).12Model (cont’d)When a broker purchases an article, it immediately sends it to a set of subscribing consumers, paying a transportation fee PT to the system.Subscribers examine the article and pay to the broker (Pb) if they want to keep the article (consume it).When a consumer receives an article, it pays PC computation fee to the system for evaluation.The broker’s decision method is parametrized through (random process).13Analysis of the SystemUtility maximization settingProbabilistic setting for modeling purposes.Problem: Maximize expected utility per article for consumers and brokers.Even for a simple model the expected profit per article formulae (for brokers and consumers) become quite involved (see the paper).Computation requires a lot of state information.In a distributed system such information would be unavailable/prohibitively expensive to obtain.14System with one brokerA single broker offers a single category to an arbitrary large number of consumers.B=1, CThe broker and the consumers have complete info about the system.They act instantaneously to maximize their profit:Broker chooses a set of preferred consumers;Consumers decide whether they want to subscribe to the broker or not.The broker and the consumers agree that the subscription makes sense if consumer’s interest is above a certain threshold.They disagree on what the threshold should be.If both sides can reach an agreement an equilibrium price is established.15Two-broker SystemAdd another broker and another information category.Each broker sets its price and interest level to the value that maximizes its profit given the other broker’s parameters;Inter-agent feedback.The settings are not changed until the other broker moves its parameters.16System Behavior: Two-broker caseBrokers make adjustments at every time tick: P1(t+1)=P1*(P2(t)), P2(t+1)=P2*(P1(t)), etc. Emergent phenomena:Price war;Spontaneous specialization.Asymptotic behavior of the system is a limit cycle:Regardless of the initial price p1, two brokers will become trapped in a price war.First, each broker tries to grab a large portion of the market by undercutting the price of the other broker.Eventually, losses overweigh the desire to undercut the competitor, and a broker raises the price sharply.The second broker raises the price in response.Undercutting starts again.17Price warsResults in a more complex analog of a price war.Both the prices and the set of categories offered cycle endlessly.Price wars are harmful to the brokers.Each broker gets only half of the utility it expected to get.Going for an instantaneous global optimum does not work in a system with more then one broker.Greedy strategy fails.18Can it be overcome?The agents are myopic: there decisions are made without anticipating the other player’s response.Solution: add foresight:An agent modeling another agent;Level-one agent models its
View Full Document
Unlocking...