Autonomous agent learning using an actor-critic algorithm and behavior models

We introdu e a Supervised Reinfor ement Learning (SRL) algorithm for autonomous learning problems where an agent is required to deal with high dimensional spa es. In our learning algorithm, behavior models learned from a set of examples, are used to dynami ally redu e the set of relevant a tions at ea h state of the environment en ountered by the agent. Su h subsets of a tions are used to guide the agent through promising parts of the a tion spa e, avoiding the sele tion of useless a tions. The algorithm handles ontinuous states and a tions. Our experimental work with a di ult robot learning task shows learly how this approa h an signi antly speed up the learning pro ess and improve the nal performan e.