Adding domain knowledge to a monte carlo tree search algorithm in the game of Go

Abstract

In this thesis the following research question will be looked into: "Is it viable to apply domain specific knowledge to a monte-carlo tree search go playing algorithm". After an introduction to the game of go itself, and monte carlo tree search, the domain knowledge that is used and different methods of implementing this will be explained. To answer the research question, an algorithm was made, which will be explained, followed by the results. The results of the tests that were run, show that adding domain knowledge to a go playing Monte Carlo tree search algorithm is defiantly viable. The combination of domain knowledge to help guide the learning element in MCTS, as a way of evaluating the moves that are possible, clearly has got its benefits. Although, depending on the algorithm, it is important to note that increasing the amount of simulations does not automatically mean a better score. So the balance between the variables involved in the MCTS algorithm, and the knowledge of go added into the algorithm is needed when looking for the best results.