A machine learning application to computer-aided engineering

Concerns the automatic provision of knowledge-based systems in computer-aided engineering. A machine learning approach allows the extraction of knowledge (how to solve a problem) from examples (cases for which the solution is known); the expert then formulates the problem, and gathers or builds typical examples. To cope with the technical data, a multilayer learning algorithm is designed. Data often are insufficiently representative from a statistical point of view; a symbolic treatment provides rules adapted to the general case, from prototypical cases given by the experts (experts consider particular domains such as spheres or cracks). As the problem is still theoretically unsolved, the optimality criterion first regards the predictive accuracy of the rule set. Multilayer learning achieves a resolution through successive symbolic approximations: a set of approximate rules is provided in one learning step, the next learning step achieves the refinement of previous rules, and so on. An application of this approach to numerical analysis of structures is described. The aim is to provide an a priori error estimate for resolution by finite element methods in elastoplasticity