There is No Free Lunch: Tradeoffs in the Utility of Learned Knowledge

摘要

With the recent introduction of learning in integrated systems, there is a need to measure the utility of learned knowledge for these more complex systems. A difficulty arrises when there are multiple, possibly conflicting, utility metrics to be measured. In this paper, we present schemes which trade off conflicting utility metrics in order to achieve some global performance objectives. In particular, we present a case study of a multi-strategy machine learning system, mutual theory refinement, which refines world models for an integrated reactive system, the Entropy Reduction Engine. We provide experimental results on the utility of learned knowledge in two conflicting metrics - improved accuracy and degraded efficiency. We then demonstrate two ways to trade off these metrics. In each, some learned knowledge is either approximated or dynamically 'forgotten' so as to improve efficiency while degrading accuracy only slightly.