Genetic algorithms applied to formal neural networks: parallel genetic implementation of a Boltzmann machine and associated robotic experimentations
Pierre Bessìère · HAL (Le Centre pour la Communication Scientifique Directe) · 1991
In this paper we describe a possible application of computing techniques inspired by natural life mechanisms (genetic algorithms and artificial neural networks) to an artificial life creature, namely a small mobile robot, called KitBorg. We proposed in a previous work (Bessière 1990) Probabilistic Inference as a possible underlying theory or mathematical metaphor for numerous works in the field of formal neural networks. Probabilistic Inference suggests that any cognitive problem may be split in two optimization problems. The first one called "dynamic inference problem" is an abstraction of "learning", the second one, namely, the "static inference problem", being a mathematical metaphor of "pattern association". In this previous paper, for instance, Boltzmann machines have been shown to be a special case of probabilistic inference, where the two optimization problems are dealt with using simulated annealing (Kirckpatrick 1983) for the pattern association part and using simple gradient descent for the learning one. It was, then, suggested that other optimization technics should be considered in that context and especially genetic algorithms. The purpose of this paper is to describe the state of the art of the investigations we are making about that question using a parallel genetic algorithm. We will first recall the principles of probabilistic inference, then , we will present briefly the parallel genetic algorithm and the ways it is used to deal with both optimization problems, to finally conclude about ongoing robotic experimentations and future planned extensions.