A Defense Model for Games with Incomplete Information

Abstract

Making a decision, an agent must consider how his outcome can be influenced by possible actions of other agents. A’b est defense model’ for games involving uncertainty assumes usually that the opponents know everything about the actual situation and the player’s plans for certain. In this paper it’s argued that the assumption results in algorithms that are too cautious to be good in many game settings. Instead, a ’reasonably good defense’ model is proposed: the player should look for a best strategy against all the potential actions of the opponents, still assuming that any opponent plays his best according to his actual knowledge. The defense model is formalized for the case of two-player zero-sum (adversary) games. Also, algorithms for decision-making against ’reasonably good defense’ are proposed.The argument and the ideas are supported by the results of experiments with random zero-sum two-player games on binary trees.KeywordsMonte CarloIncomplete InformationMultiagent SystemGood DefenseGame TreeThese keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.