Bounding Performance Loss in Approximate MDP Homomorphisms

Part of Advances in Neural Information Processing Systems 21 (NIPS 2008)

Bibtex Metadata Paper


Jonathan Taylor, Doina Precup, Prakash Panagaden


We define a metric for measuring behavior similarity between states in a Markov decision process (MDP), in which action similarity is taken into account. We show that the kernel of our metric corresponds exactly to the classes of states defined by MDP homomorphisms (Ravindran \& Barto, 2003). We prove that the difference in the optimal value function of different states can be upper-bounded by the value of this metric, and that the bound is tighter than that provided by bisimulation metrics (Ferns et al. 2004, 2005). Our results hold both for discrete and for continuous actions. We provide an algorithm for constructing approximate homomorphisms, by using this metric to identify states that can be grouped together, as well as actions that can be matched. Previous research on this topic is based mainly on heuristics.