ABALEARN: A PROGRAM THAT LEARNS HOW TO PLAY ABALONE

Abstract

The article describes ABALEARN, a self-learning Abalone program capable of automatically reaching an intermediate level of play. This is achieved without expert-labelled training examples, deep searches, and exposure to competent play.

Our approach is based on a reinforcement-learning algorithm that is risk seeking, since defensive players in Abalone tend to postpone a game endlessly. We show that risk sensitivity allows a successful self-play training. Moreover, we propose a set of features that seem relevant for achieving a rather skilled level of play.

We evaluate our approach using a fixed heuristic opponent as a benchmark. We pit our agents against human players on-line and compare samples of our agents at different times of training.