Improved upper bounds on the expected error in constant step-size Q-learning
Carolyn L. Beck, R. Srikant · 2013
We consider fixed step-size Q-learning algorithms applied to finite state and action space, discounted reward Markov decision problems (MDPs). In previous work we derived a bound on the first moment of the Q-value estimation error, specifically on the expected steady-state value of the infinity norm of the error. The goal in both this paper, and the previous, is to maximize a discounted sum of rewards over an infinite time horizon. However, in our previous work, the bound we derived holds only when the step-size is sufficiently, and sometimes impractically, small. In this paper, we present a new error bound that, as before, goes to zero as the step-size goes to zero, but is also valid for all values of the step-size. To obtain the new bound, we divide time into frames such that the probability that there is some state that is not visited within the frame is strictly less than 1: Our error bound is then found by sampling the system one time in every frame.