Reinforcement learning from comparisons: Three alternatives is enough, two is not

Abstract

The paper deals with the problem of finding the best alternatives on the basis of pairwise comparisons when these comparisons need not be transitive. In this setting, we study a reinforcement urn model. We prove convergence to the optimal solution when reinforcement of a winning alternative occurs each time after considering three random alternatives. The simpler process, which reinforces the winner of a random pair does not always converges: it may cycle.