On Synchronous Binary Log-Linear Learning and Second Order Q-learning

Abstract

The focus of this paper is on the enhancement of Log Linear Learning (LLL) and Q-learning (QL) in game theory and their applications in multi-robot control. We first propose a modified Binary Log-Linear Learning (BLLL) algorithm that can achieve a better performance and higher learning rate when is compared to standard BLLL. However, due to a number of assumptions, practical applicability of a LLL-based algorithm is limited. To relax this limitation we then propose a modified QL algorithm that can achieve the same performance but with the price of lower learning rate. The algorithms proposed are tested numerically.