On the Job Training

Abstract

We propose a new framework for building and evaluating machine learning algorithms. We argue that many real-world problems require an agent which must quickly learn to respond to demands, yet can continue to perform and respond to new training throughout its useful life. We give a framework for how such agents can be built, describe several metrics for evaluating them, and show that subtle changes in system construction can significantly affect agent performance.

8 pages, submitted to NIPS 2005