Towards AI Robustness Multi-Agent Adversarial Planning in Game Play

This paper applied Monte Carlo Tree Search (MCTS) and its variant algorithm “Upper Confidence bounds applied to Trees” (UCT) to Chinese checker game and visualized it in the Ludii games portal. Since the MCTS approach cannot guarantee a finite game length in our simulation and the winning strategy is inefficient, we developed a Convolution Neural Network (CNN) model formulated on top of the MCTS algorithm to improve the performance of the Chinese checker game. Our experiment and simulations show that by using the weak labels generated by MCTS algorithm, the CNN-based model could learn without supervised signals from human interventions and execute the game strategy in a finite time. PyGame and the Ludii game portal are used in our simulation to visualize the game and show the final game results.