Learning the suitability of simple behaviors to obtain composite behaviors for autonomous agents♣

The composition of simple behaviors is a common practice to obtain complex behaviors from autonomous agents. In this paper, we present S-ELF, a reinforcement learning approach that learns to coordinate pre-defined basic behaviors and is derived from ELF (Evolutonary Learning of Fuzzy rules) (Bonarini, 1993), (Bonarini, 1996a). S-ELF (Symbolic ELF) learns the context of activation for each of the basic behaviors available. It works with contexts described by logical expressions composed by conjunctions of both positive and negative, high-level, fuzzy predicates, such as: "in corridor-1" or "face door-2". S_ELF generates general, fuzzy rules that coordinate the activation of basic behaviors. S-ELF has been tested in a framework similar to that of Flakey (Saffiotti et al., 1995), where navigation was programmed in terms of behaviors and their context of activation. Learning a control system described by high-level predicates brings to robust behaviors that can be instantiated in different environments, by adapting the meaning of the used predicates. In the paper, we first present the architecture of the fuzzy control system that we have adopted. Then, we describe the main features of S-ELF, focusing on some learning mechanisms that may be interesting for any reinforcement learning algorithm. Finally, we briefly mention the experimental results that we have obtained.