Continuous Rapid Action Value Estimates

Adrien Couetoux (1) , Mario Milone (1) , Matyas Brendel (2) , Hassen Doghmen (2) , Michèle Sebag (2, 1) , Olivier Teytaud (1, 2)

1 LRI - Laboratoire de Recherche en Informatique
2 TAO - Machine Learning and Optimisation

In the last decade, Monte-Carlo Tree Search (MCTS) has revolutionized the domain of large-scale Markov Decision Process problems. MCTS most often uses the Upper Confidence Tree algorithm to handle the exploration versus exploitation trade-off, while a few heuristics are used to guide the exploration in large search spaces. Among these heuristics is Rapid Action Value Estimate (RAVE). This paper is concerned with extending the RAVE heuristics to continuous action and state spaces. The approach is experimentally validated on two artificial benchmark problems: the treasure hunt game, and a real-world energy management problem.

Mots clés: Rapid Action-Value Estimates; continuous domains; reinforcement learning

Domaines: Optimisation et contrôle [math.OC]

Adrien Couetoux, Mario Milone, Matyas Brendel, Hassen Doghmen, Michèle Sebag, et al.. Continuous Rapid Action Value Estimates. The 3rd Asian Conference on Machine Learning (ACML2011), Nov 2011, Taoyuan, Taiwan. pp.19-31. ⟨hal-00642459⟩