Match-3 game has simple game-play but it is hard for AI to learning specific strategies. Here, we present our self designed match-3 game bubble crush. The game mimics the situation that players make moves to earn scores to win different levels on the Unity and ML-agent platform. Traditional studies in Match 3 games have attempted to utilize Monte Carlo tree search (MCTS) and convolutional neural networks (CNNs), but this articles demonstrated how Reinforcement Learning (RL) method based on Proximal Policy Optimization(PPO) outperforms random policy AI and provides metrics for evaluation of agents and corresponding baselines in different scenarios.
Our EDD, paper, presentation and video demo are in the Bubble Crush Google Drive.