This paper describes a new adaptive Monte Carlo Tree Search (MCTS) algorithm that uses evolution to rapidly optimise its performance. An evolutionary algorithm is used as a source of control parameters to modify the behaviour of each iteration (i.e. each simulation or roll-out) of the MCTS algorithm; in this paper we largely restrict this to modifying the behaviour of the random default policy, though it can also be applied to modify the tree policy. This method of tightly integrating evolution into the MCTS algorithm means that evolutionary adaptation occurs on a much faster time-scale than has previously been achieved, and addresses a particular problem with MCTS which frequently occurs in real-time video and control problems: that uniform random roll-outs may be uninformative. Results are presented on the classic Mountain Car reinforcement learning benchmark and also on a simplified version of Space Invaders. The results clearly demonstrate the value of the approach, significantly outperforming "standard" MCTS in each case. Furthermore, the adaptation is almost immediate, with no perceptual delay as the system learns: the agent frequently performs well from its very first game.
展开▼
