An Introduction to Temporal Difference Learning

Temporal Difference learning is one of the most used approaches for policy evaluation. It is a central part of solving reinforcement learning tasks. For deriving optimal control, policies have to be evaluated. This task requires value function approximation. At this point TD methods find application. The use of eligibility traces for backpropagation of updates as well as the bootstrapping of the prediction for every update state make these methods so powerful. This paper gives an introduction to reinforcement learning for a novice to understand the TD(λ) algorithm as presented by R. Sutton. The TD methods are the center of this paper, and hence, each step for deriving the update function is treated. Starting with value function approximation followed by the Bellman Equation and ending with eligibility traces. The further enhancement of TD in form of linear-squared temporal difference methods is treated. Both methods are compared in respect to their computational cost and learning rate. In the end an outlook towards application in control is given.