A Comparative Study of Opponent Type Effects on Speed of Learning for an Adversarial Q-Learning Agent

One goal of artificial intelligence has been to approximate human thought. The ability to learn is common to both humans and artificial intelligence agents, but the methods can sometimes differ. Reinforcement learning is one area where machine learning was initially based on the learning methods of humans and animals. In the case of two-player, turn-based adversarial games, it is advantageous to be able to train the learning agent against an opponent. The skill level of the opponent, however, is important, and can lead to different learning speeds, breadth of learning, or efficiency. We demonstrate the outcomes when having a learning agent play a simple game (tic-tac-toe) against an unskilled opponent, an opponent of equal skill, and an advanced opponent.