S. Whisenhunt and D. J. Cook
The focus of this project is to develop methodologies for using machine learning techniques in adversarial robot situations. In particular, we are using multiple robots to play a version of the wumpus world game. In this game, one robot represents the agent and a second robot represents the wumpus. Our goal is for the agent robot to make autonomous decisions that allow it to elude the wumpus, grab the gold and win the game. To achieve this goal, we consult several supervised machine learning algorithms to decide the agent’s move. Agent moves are learned from training examples encoding characteristics of the world, the game state, and the predicted wumpus move. In this paper we will compare the performance of a decision tree learner, a naive Bayesian classifier, a backpropagation neural network, and a learning-based belief network on actual wumpus world games.