Semantic Networks

For the learning phase, Levinson used a combination of rote learning as in case-based reasoning, restructuring to derive significant generalizations, a similarity measure based on the generalizations, and a method of back propagation to estimate the value of any case that occurred in a game. For playing chess, the cases were board positions represented as graphs. Every position that occurred in a game was stored in a generalization hierarchy, such as those used in definitional networks. At the end of each game, the system used back propagation to adjust the estimated values of each position that led to the win, loss, or draw. When playing a game, the system would examine all legal moves from a given position, search for similar positions in the hierarchy, and choose the move that led to a position whose closest match had the best predicted value.