Evaluating a reinforcement learning algorithm with a general intelligence test

Abstract

In this paper we apply the recent notion of anytime universal intelligence tests to the evaluation of a popular reinforcement learning algorithm, Q-learning. We show that a general approach to intelligence evaluation of AI algorithms is feasible. This top-down (theory-derived) approach is based on a generation of environments under a Solomonoff universal distribution instead of using a pre-defined set of specific tasks, such as mazes, problem repositories, etc. This first application of a general intelligence test to a reinforcement learning algorithm brings us to the issue of task-specific vs. general AI agents. This, in turn, suggests new avenues for AI agent evaluation and AI competitions, and also conveys some further insights about the performance of specific algorithms. © 2011 Springer-Verlag.