A computational intelligence algorithm for simulation-driven optimization problems

The modern engineering design process often relies on computer simulations to evaluate candidate designs. This simulation-driven approach results in what is commonly termed a computationally expensive black-box optimization problem. In practise, there will often exist candidate designs which cause the simulation to fail. Such simulation failures can consume a large portion of the allotted computational resources, and thus can lead to search stagnation and a poor final solution. To address this issue, this study proposes a new computational intelligence optimization algorithm which combines a model and a k-NN classifier. The latter predicts which solutions are expected to cause the simulation to fail, and its prediction is incorporated with the model prediction to bias the search towards valid solutions, namely, for which the simulation is expected to succeed. A main contribution of this study is that to further improve the search efficacy, the proposed algorithm leverages on model-selection theory and continuously calibrates the classifier during the search. An extensive performance analysis using an engineering application of airfoil shape optimization shows the efficacy of the proposed algorithm.