A Reinforcement Learning Approach for a Goal-Reaching Behavior
Tiziana D’Orazio, Grazia Cicirelli, Giovanni Attolico, Cosimo Distante · The Florida AI Research Society · 1999
Developing elementary behavior is the starting point for the realization of complex systems. In this paper we win describe a learning algorithm that realizes a simple goal-reaching behavior for an autonomous vehicle when a-priori knowledge of the environment is not provided. The state of the system is based on information received by a visual sensor. A Q-learnlng algorithm associates the optimal action to each state, developing the optimal state-action rules (optimal policy). A few training trials are sufficient, in simulation, to learn she optimal policy since during the test trials the set of actions is initially limited. The state and action sets are then enlarged, introducing fuzzy variables with their membership functions to the extent of tackling errors in state estimation due to the noise in the vision measurements. Experimental resuILs, both in simulated and real environment, are shown.