Reinforcement Learning: Insights from Interesting Failures in Parameter Selection

Abstract

We investigate reinforcement learning methods, namely the temporal difference learning TD(λ) algorithm, on game-learning tasks. Small modifications in algorithm setup and parameter choice can have significant impact on success or failure to learn. We demonstrate that small differences in input features influence significantly the learning process. By selecting the right feature set we found good results within only 1/100 of the learning steps reported in the literature. Different metrics for measuring success in a reproducible manner are developed. We discuss why linear output functions are often preferable compared to sigmoid output functions.