WallZero

First-author research — Hsing-Yu Chen, Jerome Arjonilla, I-Chen Wu, Ti-Rong Wu, Computers and Games (CG 2026). Open-source at rlglab/wallzero.

WallZero is an AlphaZero-based game AI that masters WallGo, a two-player territory board game. It is my first-author research, accepted at Computers and Games (CG 2026) and open-sourced by the RLG Lab.

The game — WallGo

WallGo is played on a 7×7 board where each player controls four stones. On each turn a player moves a stone and then builds a wall, progressively partitioning the board; once no more moves are possible, the player enclosing the larger territory wins. WallZero handles two settings — an empty mode that starts from a blank board, and a 4-stone mode with preset opening stones.

Method

WallZero is built on the MiniZero framework and follows the AlphaZero paradigm: a deep neural network coupled with Monte Carlo Tree Search (MCTS), trained purely through self-play with no human game records. Because WallGo combines piece movement with wall placement, the agent has to learn both fine-grained spatial control and long-term territory strategy entirely on its own. The engine and search are implemented in C++ for speed, with Python for training and tooling.

Results

• Learns WallGo from scratch via self-play reinforcement learning — no human data.
• Defeated professional Go players (3-dan and 9-dan), taking on average 1.98× more territory than its human opponents.
• Accepted as first-author work at Computers and Games (CG 2026).