LEARNING BY BUILDING CAUSAL EXPLANATIONS.

In AI systems where the task is to predict future events in a complex domain, prediction failures provide the initial focus for learning mechanisms. When a failure occurs, causal knowledge is brought to bear on the relevant data in order to construct an explanation of the failure. Causal knowledge includes knowledge of the relationships between the fundamental events and objects in a domain. The explanation built using this knowledge serves as the foundation for creation of new memory structures which prevent the failure from recurring. Two case studies provide examples of the use of causal knowledge in learning systems: HANDICAPPER, a program which predicts the outcome of horse races, and FORECASTER, a program which predicts the weather.