Extensive-Form Argumentation Games.pdfVIP

Extensive-Form Argumentation Games Ariel D. Procaccia Jeffrey S. Rosenschein School of Engineering and Computer Science The Hebrew University of Jerusalem Jerusalem, Israel Abstract Two prevalent approaches to automated negotiation are the application of game-theoretic notions and the argumentation-based angle; these two schemes are frequently at odds. An ele- gant view of argumentation is Dung’s abstract argumentation theory [2], which cold-shoulders the internal structure of arguments in favor of the entire debate’s global structure. Dung’s theory is elaborated by work in dialectical argumentation theory, which focuses on dialogues between two players. In this paper, we enhance the abovementioned frameworks by considering two-agent settings where each of the agents is identified with a set of arguments. A binary attack relation between arguments is given, as well as (most importantly) a payoff function that assigns real values to ev- ery possible valid dialogue. Such game-based argumentation frameworks can be lucidly realized as games in extensive form (marrying, in a sense, the game-theoretic and argumentation-based approaches). We investigate specialized notions of “maximally-defendable” sets of arguments, and the correlation between the properties of the agents’ argument sets and the size of the associated game tree. Furthermore, algorithmic issues are considered: we present simplifica- tions and efficient solutions for our argumentation games, and discuss a special case where the representation of the framework is logarithmic in the size of the associated game tree. 1 Introduction Negotiation between agents has emerged in recent years as a method of resolving scenarios where agents have conflicting interests but may benefit by cooperating. The work on automated nego- tiation can be classified into three general approaches [13]: game-theoretic [10, 14, 15], heuristic- based [4, 5, 9], and argumentation-based [11, 12, 16]. A major advantage of argumentation-base