Temporal Abstraction in Monte Carlo Tree Search
Mostafa Vafadost · ERA: Education and Research Archive (University of Alberta) · 2013
Given nothing but the generative model of the environment, Monte Carlo Tree Search techniques have recently shown spectacular results on domains previously thought to be intractable. In this thesis we try to develop generic techniques for temporal abstraction inside MCTS that would allow the effective construction of medium/long term plans in arbitrary Atari 2600 games. We introduce two methods: 1- CTW-UCT, that uses CTW algorithm to extract information from expert trajectories in order to make use of this information in search. 2- Variable Time Scale (VTS), that finds the desired time scale for taking each action online. We found that, CTW-UCT did not achieve satisfactory results. The VTS algorithm, however, was shown to be a promising algorithm. Although VTS did not achieve the highest individual scores on any game, it performed close to the best on most of the games. In other words, compared to different UCT algorithms that make use of macro actions with repeated actions of pre-specified lengths, VTS achieved comparable results to the best score on most of the games. We conclude therefore that, for domains where we do not know the desired time scale, VTS can be a good option.