A Classifier-Assisted Framework for Expensive Optimization Problems: A Knowledge-Mining Approach

Abstract

Real-world engineering design optimization problems often rely on computationally-expensive simulations to replace laboratory experiments. A common optimization approach is to approximate the expensive simulation with a computationally cheaper model resulting in a model-assisted optimization algorithm. A prevalent issue in such optimization problems is that the simulation may crash for some input vectors, a scenario which increases the optimization difficulty and results in wasted computer resources. While a common approach to handle such vectors is to assign them a penalized fitness and incorporate them in the model training set this can result in severe model deformation and degrade the optimization efficacy. As an alternative we propose a classifier-assisted framework where a classifier is incorporated into the optimization search and biases the optimizer away from vectors predicted to crash to simulator and with no model deformation. Performance analysis shows the proposed framework improves performance with respect to the penalty approach and that it may be possible to ‘knowledge-mine’ the classifier as a post-optimization stage to gain new insights into the problem being solved.