Relativized hierarchical decomposition of Markov decision processes.

Reinforcement Learning (RL) is a popular paradigm for sequential decision making under uncertainty. A typical RL algorithm operates with only limited knowledge of the environment and with limited feedback on the quality of the decisions. To operate effectively in complex environments, learning agents require the ability to form useful abstractions, that is, the ability to selectively ignore irrelevant details. It is difficult to derive a single representation that is useful for a large problem setting. In this chapter, we describe a hierarchical RL framework that incorporates an algebraic framework for modeling task-specific abstraction. The basic notion that we will explore is that of a homomorphism of a Markov Decision Process (MDP). We mention various extensions of the basic MDP homomorphism framework in order to accommodate different commonly understood notions of abstraction, namely, aspects of selective attention. Parts of the work described in this chapter have been reported earlier in several papers (Narayanmurthy and Ravindran, 2007, 2008; Ravindran and Barto, 2002, 2003a,b; Ravindran et al., 2007).