Heuristic Real-Time Dynamic Programming (Preliminary Results)
Blai Bonet, Hdctor Geffner · 1997
Building on the work of Korf on real-time search, and of Barto, Bradtke and Singh on real-time dy-namic programming, we discuss three features that we have found crucial for the solution of large MDP’s; namely, selective updates, selective lookahead, and good heuristic functions. We also present an RTDP algorithm that exhibits these features and assess its performance over three types of problems.