Maintaining the Utility of Learned Knowledge Using Model-Based Control

The overfit problem in empirical learning and the utility problem in analytical learning both describe a common behavior of machine learning methods: the eventual degradation of performance due to increasing amounts of learned knowledge. Plotting the performance of the changing knowledge during execution of a machine learning method (the performance response) reveals similar curves for several methods. The performance response generally indicates a single peak performance greater than that attained by popular pruning techniques. The similarity in performance responses suggests a parameterized model relating performance to the amount of learned knowledge. Given this model, a model-based adaptive control (MBAC) approach can be used to update the model based on feedback from the performance element and make control decisions regarding the amount of knowledge to be learned or unlearned. In view of the large number of alternative learning methods, a more general utility problem exists in determining not only the correct amount of learned knowledge, but also the correct method for learning this knowledge. Relying too heavily on one particular learning method may result in less than optimal performance achievement. Overcoming this general utility problem requires a new control mechanism for determining the correct learning method and amount of learned knowledge in order to achieve the performance objectives of the task. Maintaining models for several learning methods allows the MBAC approach to decide the appropriate type of learning, in addition to the amount