An Architecture for Representing and Learning Behaviors by Trial and Error
Pascal Blanchet · The MIT Press eBooks · 1994
We propose an architecture for representing and learning behaviors by trial and error. This architecture is a network of automata which is dynamically built while the system evolves in its environment. It is presented as an alternative to rule based systems which learn through temporal differences algorithms. This system learns sequences of perceptions instead of perception-action rules. These sequences represent rules of the environment, and can also be considered as production rules. The learned sequences are used for learning other sequences. An induction mechanisms is applied to produce general sequences from the different trials of the system. Different experiments has been carried out, including a block manipulation task which is described at the end of this paper. 1. Introduction The question we are interested in, is the one of learning through interaction with the external world in a way that is as close as possible to human's one. In this paper, we focus on the following pro...