Unifying Empirical and Explanation-Based Learning by Modeling the Utility of Learned Knowledge

The overrt problem in empirical learning and the utility problem in explanation-based learning describe a similar phenomenon: the degradation of performance due to an increase in the amount of learned knowledge. Plotting the performance of learned knowledge during the course of learning (the performance response) reveals a common trend for several learning methods. Modeling this trend allows a control system to constrain the amount of learned knowledge to achieve peak performance and avoid the general utility problem. Experiments evaluate a particular empirical model of the trend, and analysis of the learners derive several formal models. If, as evidence suggests, the general utility problem can be modeled using the same mechanisms for diierent learning paradigms, then the model serves to unify the paradigms into one framework capable of comparing and selecting diierent learning methods based on predicted achievable performance.