Neural Networks and Evolutionary Algorithms for Solving Multi Objective Optimisation Problems

摘要

Many of multi objective optimisation problems are often subject to parameters with uncertainties and noise. In this case, for identify the robust solutions we add generally small amounts of noise and evaluate it with Monte Carlo simulation. In this paper, we propose a new methodology for solving these types of multi objective optimisation problems. This methodology consists to augment the objective functions space with robustness functions in order to find a compromise between robustness and optimality. The multiobjective optimisation problem is solved with an evolutionary algorithm. A neural network is used to reduce considerably the computing time, in particular for the robustness function evaluations. The space of optimal solutions is of high dimension. In practice, it is difficult to exploit usefully these solutions. One approach would be to use a clustering technique (datamining for example). One of popular datarmining techniques is the Self Organizing Map (SOM) proposed by Kohonen. These techniques make it possible to find all the clusters in the cost functions and design space and then it constitute a powerful tools of design decision-making Numerical example illustrates the interest and the performance of the suggested method in a multi objective optimisation of stochastic dynamics.