Handling Ambiguous Effects in Action Learning

Boris Lesner (1) , Bruno Zanuttini (2)

1 GREYC - Groupe de Recherche en Informatique, Image et Instrumentation de Caen
2 Equipe MAD - Laboratoire GREYC - UMR6072

We study the problem of learning stochastic actions in propositional, factored environments, and precisely the problem of identifying STRIPS-like effects from transitions in which they are ambiguous. We give an unbiased, maximum likelihood approach, and show that maximally likely actions can be computed efficiently from observations. We also discuss how this study can be used to extend an RL approach for actions with independent effects to one for actions with correlated effects.

Domaines: Intelligence artificielle [cs.AI]

Boris Lesner, Bruno Zanuttini. Handling Ambiguous Effects in Action Learning. 9th European Workshop on Reinforcement Learning (EWRL 2011), 2011, Greece. 12 p. ⟨hal-00946967⟩