An actor-critic method using Least Squares Temporal Difference learning

In this paper, we use a Least Squares Temporal Difference (LSTD) algorithm in an actor-critic framework where the actor and the critic operate concurrently. That is, instead of learning the value function or policy gradient of a fixed policy, the critic carries out its learning on one sample path while the policy is slowly varying. Convergence of such a process has previously been proven for the first order TD algorithms, TD(¿) and TD(1). However, the conversion to the more powerful LSTD turns out not straightforward, because some conditions on the stepsize sequences must be modified for the LSTD case. We propose a solution and prove the convergence of the process. Furthermore, we apply the LSTD actor-critic to an application of intelligently dispatching forklifts in a warehouse.