Approximate Solutions To Constrained Risk-Sensitive Markov Decision Processes

TL;DR

This paper develops methods to find near-optimal, feasible policies for constrained risk-sensitive Markov decision processes, using finite-horizon approximations and perturbations to handle infinite-horizon problems.

Contribution

It introduces two approximation techniques for solving constrained risk-sensitive MDPs, ensuring near-optimality and feasibility through finite-horizon models.

Findings

Existence of solutions if the problem is feasible

Two methods for approximate policy computation

Bounded constraint violations for near-optimal policies

Abstract

This paper considers the problem of finding near-optimal Markovian randomized (MR) policies for finite-state-action, infinite-horizon, constrained risk-sensitive Markov decision processes (CRSMDPs). Constraints are in the form of standard expected discounted cost functions as well as expected risk-sensitive discounted cost functions over finite and infinite horizons. The main contribution is to show that the problem possesses a solution if it is feasible, and to provide two methods for finding an approximate solution in the form of an ultimately stationary (US) MR policy. The latter is achieved through two approximating finite-horizon CRSMDPs which are constructed from the original CRSMDP by time-truncating the original objective and constraint cost functions, and suitably perturbing the constraint upper bounds. The first approximation gives a US policy which is $\epsilon$-optimal and feasible for the original problem, while the second approximation gives a near-optimal US policy whose violation of the original constraints is bounded above by a specified $\epsilon$. A key step in the proofs is an appropriate choice of a metric that makes the set of infinite-horizon MR policies and the feasible regions of the three CRSMDPs compact, and the objective and constraint functions continuous. A linear-programming-based formulation for solving the approximating finite-horizon CRSMDPs is also given.