- Abstract: Temporal Difference learning with function approximation has been widely used recently and has led to several successful results. However, compared with the original tabular-based methods, one major drawback of temporal difference learning with neural networks and other function approximators is that they tend to over-generalize across temporally successive states, resulting in slow convergence and even instability. In this work, we propose a novel TD learning method, Hadamard product Regularized TD (HR-TD), that reduces over-generalization and thus leads to faster convergence. This approach can be easily applied to both linear and nonlinear function approximators. HR-TD is evaluated on several linear and nonlinear benchmark domains, where we show improvement in learning behavior and performance.
- Keywords: Reinforcement Learning, TD Learning, Deep Learning
- TL;DR: A regularization technique for TD learning that avoids temporal over-generalization, especially in Deep Networks