An Application of Neural Networks to a Non-Deterministic Game of Imperfect Information

Abstract

Current research into game-playing neural networks emphasizes deterministic games of per- fect information and training with “expert knowledge” such as known successful strategies. As an example, Google’s AlphaGo, a neural network agent designed to play the game Go, was trained with over 30 million different board positions drawn from approximately 160,000 different games. This expert knowledge gives AlphaGo a solid baseline for future play. We studied the application of neural networks to the game Lost Cities, a non-deterministic game of imperfect information, without providing such expert knowledge, and designed several neural network agents and implemented an agent using Monte Carlo tree search for comparison.