Monte Carlo algorithms for time-constrained general game playing

Abstract

General Game Playing (GGP) is a complex field for Artificial Intelligence (AI) agents because it demands the ability to play varied games without prior knowledge. This paper introduces two algorithms to enhance move suggestions in time-limited GGP. Our first strategy is a modification of Sequential Halving Applied to Trees (SHOT), a non-exploiting algorithm. The second strategy is a hybrid version of Upper Confidence Tree (UCT) that combines Sequential Halving and UCB√ to focus more on acquiring information at the root node. To test how agents perform, we use three diferente evaluation scenarios. First, we observe how resources are allocated among the selection policies. Next, we compare the performance of these strategies over five different board games with a set number of playouts, and in a competitive GGP environment where each game is played in one minute. These tests allow us to analyze the outcomes and implications of our proposed strategies.