Using SARSA with function approximation to create policies for MCTS

Abstract

This paper proposes CARL, a pair of agents that apply reinforcement learning and function approximation using regression to learn policies for games where human heuristics cannot be applied. The purpose of these policies is to do search control in Monte Carlo Tree Search (MCTS), a heuristic search algorithm to see if the learned policies can outperform upper confidence bound for trees (UCT).