Games in AI are always the ideal test bed for AI research, due to their clear results, motivation, and challenge. The Malmo platform is a sophisticated AI experimentation platform built on top of Minecraft, and designed to support fundamental research in artificial intelligence. The Malmo collaborative AI includes the following challenges: (1) Interactions among agents, (2) Uncertainty, (3) Sequential decision making, and (4) Real-time decision making. The Catch-the-Pig Game is one of the famous collaborative games. It is both competitive and collaborative. While one agent wants to collaborate, he also needs to worry about the other agent betraying him. There are also some unknown parameters that need to be learned for both players during the game.
Dr. An introduced his previous work on applying adaptive Q-learning to the Catch-the-Pig Game. He described several challenges and the corresponding solutions while designing the AI of the game. They used Bayesian update of agent types to conquer the uncertainty, and human-reasoning initialization to deal with the slow start problem of the training period. State abstraction also helps to reduce the problem size where huge problem size is a common problem for complicated game configuration. The experiments also tell us that simply learning a policy by ignoring the other agent’s type does not work, which implies we should not only focus on the profit of ourselves but the interest of our adversary or collaborator. Future research includes completely unknown collaborators and online setting of the games.
