From Optimization to Control: Quasi Policy Iteration
This work addresses control algorithm efficiency for researchers and practitioners in reinforcement learning, but it is incremental as it builds on existing optimization and policy iteration methods.
The paper tackled the problem of designing efficient control algorithms for Markov decision processes by introducing quasi-policy iteration (QPI), which adapts the quasi-Newton method from optimization to approximate the Hessian matrix in policy iteration, resulting in empirical convergence similar to quasi-Newton methods with low sensitivity to the discount factor.
Recent control algorithms for Markov decision processes (MDPs) have been designed using an implicit analogy with well-established optimization algorithms. In this paper, we adopt the quasi-Newton method (QNM) from convex optimization to introduce a novel control algorithm coined as quasi-policy iteration (QPI). In particular, QPI is based on a novel approximation of the ``Hessian'' matrix in the policy iteration algorithm, which exploits two linear structural constraints specific to MDPs and allows for the incorporation of prior information on the transition probability kernel. While the proposed algorithm has the same computational complexity as value iteration, it exhibits an empirical convergence behavior similar to that of QNM with a low sensitivity to the discount factor.