Abstract
We present two novel approaches to accelerate the convergence of a class of Markov decision problems (MDPs) with stationary state space components, which iterate on reduced state and action spaces and converge to an optimal policy. The time aggregation based algorithm (TABA) partitions the state space into disjoint sets, one for each possible combination of the non-stationary components, and iterates in the sets. The local search policy set iteration (LSPSI) algorithm introduces a novel modified policy evaluation procedure that seamlessly performs a local search over a very large set of candidate policies, by sampling a reduced subset of actions at each state's value function update. Both approaches, as well as a combination of them, are validated by means of a mining supply chain application with a large number of stationary state components and a large set of feasible actions. The experiments suggest that the proposed frameworks are very efficient for such a class of problems.

This publication has 32 references indexed in Scilit: