Abstract
We introduce our competitive implementation of a Monte Carlo Tree Search (MCTS) algorithm for the board game of Go: Pachi. The software is based both on previously published methods and our original improvements.
We then focus on improving the tree search performance by collecting information regarding tactical situations and game status from the Monte Carlo simulations and sharing it with and within the game tree. We propose specific methods of such sharing - dynamic komi, criticality-based biasing, and liberty maps - and demonstrate their positive effect based on collected play-testing measurements.
We also outline some promising future research directions related to our work.