Symbolic Noise Detection in the Noisy Iterated Chicken Game and the Noisy Iterated Battle of the Sexes

Tsz-Chiu Au, Sarit Kraus, Dana Nau

Symbolic noise detection (SND) has been shown to be highly effective in the Noisy Iterated Prisoner’s Dilemma, in which an action can accidentally be changed into a different action. This paper evaluates this technique in two other 2 x 2 repeated games: the Noisy Iterated Chicken Game (ICG) and the Noisy Iterated Battle of the Sexes (IBS). We present a generalization of SND that can be wrapped around any existing strategy. To test its performance, we organized ICG and IBS tournaments in which we solicited several dozen strategies from different authors, and we tested these strategies with and without our SND wrapper. In our tests, SND identified and corrected noise with 71% accuracy in the ICG, and 59% accuracy in the IBS. We believe the reason why SND was less effective in the ICG was because of a tendency for IBS strategies to change more frequently from one pattern of interactions to another, causing SND to make a higher number of wrong corrections. This leads us to believe that SND will be more effective in any game in which strategies often show a stable behavior.

Subjects: 7.1 Multi-Agent Systems; 12.1 Reinforcement Learning

Submitted: Jun 20, 2008


This page is copyrighted by AAAI. All rights reserved. Your use of this site constitutes acceptance of all of AAAI's terms and conditions and privacy policy.