Loading...
Please wait, while we are loading the content...
No Free Lunches in Multi-Agent Systems , — a Characteristic Distribution Approach to Game Theoretic Modelling
| Content Provider | Semantic Scholar |
|---|---|
| Copyright Year | 2007 |
| Abstract | We introduce the notion of Characteristic Distributionswhich is a way of representing information about the payoffs of different behaviors in a Multi-age nt System. We discuss how they can be used to simplify and structure the analysis of strategies an d prove i) the existence of optimal environments, given a certain behavior, and ii) that all behavio rs payoff equally, when taken over all possible environments (no free lunch theorem for strategie s). 1: Introdu tion In Multi-Agent Systems, ( MAS), agents interact with each other and the environment in ord er to meet their design objectives. From a game theoretic point of view, this process is a choice of strategies for playing games i.e. choosing behavioral patt erns in a given environment, a choice that itself is a meta-game. Previous work on meta-games (the game of selecting a strategy for a game) include [1, 2, 7]. We distinguish iterated games from repeated ones. In repeat d games the players have no memory, while in the iterated games, the players remember all th e previous actions made by others, i.e. they have a history of the game so far. Agents are in general ab le to remember previous actions taken by themselves and other agents, and are thus well-suit ed to be modeled by means of iterated games [6]. Iterated strategic games are known to be harder to analyze an d find equilibria in than repeated ones because of the exponentially increasing number of poss ible states (and thus taken into consideration when choosing the next move). Recently, promising a ttempts have been made, especially in the field of evolutionary game theory, to use e.g. adaptive dynamics to describe how equilibria might be reached among simple strategies in iterated games [ 4, 8, 9]. While these attempts try to answer the question “How are strategies behaving?” , we will here try to focus on the questions “What is the result of their behavior?”and“How can this result be used at the meta-level?” . To help us answer these questions we will use Characteristic DistributionsorChDs. The paper is organized as follows. First, we describe the dis tinction between agents and strategies and cover some formalities. Then some properties of the appr o ch are discussed and exemplified. Finally some conclusions and further work are presented. 2: Chara teristi Distributions (ChDs) We assume an environment in which we make a distinction betwe en the agent level and the strategy levelin the sense that the strategies are unable to model agents or reason about them in Who will I meet? What are the payoffs? What is my optimal choice? Agent 1 Agent 2 Strat 1 S tr at 1 . . . . . . . . . . . . . . . . . . . . . . . . |
| File Format | PDF HTM / HTML |
| Alternate Webpage(s) | http://www.ipd.bth.se/~sja/publications/ICCIMA01.ps.gz |
| Language | English |
| Access Restriction | Open |
| Subject Keyword | Abnormal behavior Choose (action) Game theory Intelligent agent Iterated function Iteration Limbo No free lunch theorem Object-relational database Point of View (computer hardware company) Utility polyhedraloligosilsesquioxane |
| Content Type | Text |
| Resource Type | Article |
