An empirical analysis of reinforcement learning using design of experiments
Christopher J. Gatti, Mark J. Embrechts, Jonathan D. Linton · 2013
Abstract. This study uses a design of experiments approach to under-stand the behavior of a neural network to learn the mountain car domain using the TD(λ) algorithm. A large experiment is first performed to char-acterize the probability of empirical convergence based on three parameters of the TD(λ) algorithm (λ, γ, ), and a logistic regression model is fitted to this data. A detailed analysis of the parameter subspace finds that, upon convergence, these parameters significant affect convergence speed and mean performance, though performance differences are minimal. 1