Assessing the Convergence Properties of NSGA-II for Direct Crashworthiness Optimization

The elitist non-dominated sorting genetic algorithm (NSGA-II) converges to the Pareto optimal front (POF) if a sufficient number of function evaluations are allowed. However, for expensive problems involving crash simulations, only a limited number of simulations might be affordable. It is observed that initially there are significant advances towards the POF but as the population matures, the improvements are relatively small. This means that one can probably limit the computational expense by terminating the search at the right point. The paper also demonstrates a successful use of IBM cluster for parallel processing that significantly reduces clock time for optimization.