| The application of meta-learning requires a continuous adaptation to a dynamic environment. In the adaptation context, to help development of meta-learning algorithms, we suggest to use genetics algorithms like an adaptive system. An adaptive system of optimization of knowledge in direct interaction with the environment is thus recommended using genetic algorithms. To guarantee the success of such a prototype of learning, the concept of meta-learning must then be examined rigorously, which implies two possibilities of meta-learning: one centered on GA itself, according to its capacity to control its own operation by meta-learning and the second possibility, the use of GA's to carry out meta-learning. In this case, the learning is centered on the gained experiences of the system by accumulating meta-knowledge. These two possibilities led to the development of the following prototypes: (1) Prototype APAG from French "Auto-adaptation des Parametres de l'Algorithme Genetique", (2)Prototype MAAG from French " Meta-Apprentissage des Algorithmes Genetiques".; Prototype APAG is primarily based on the concept of autonomy. Autonomy is reflected by a special architecture of the individual of the genetic algorithm compared to the traditional individuals of GA's.; Prototype MAAG is characterized by its modulator aspect, with an important orientation on the concept of meta-knowledge. The principal module of MAAG is the Darwin Brain module, which contains a center of gravity of the learning system. The learning is based on the evolution by genetic algorithm, but also in combination with the neuronal Darwinism. (Abstract shortened by UMI.)... |
