Robustness Optimization for Constrained, Nonlinear Programming Problems

In realistic situations, engineering designs should take into consideration random aberrations from the stipulated design variables arising from manufacturing variability. Moreover, many environmental parameters are often stochastic in nature. Traditional nonlinear optimization attempts to find a deterministic optimum of a cost function and does not take into account the effect of these random variations on the objective. This paper attempts to devise a technique for finding optima of constrained nonlinear functions that are robust with respect to such variations. The expectation of the function over a domain of aberrations in the parameters is taken as a measure of `robustness' of the function value at a point. It is pointed out that robustness optimization is ideally an attempt to trade off between `optimality' and `robustness'. A newly-developed multicriteria optimization technique, known as Normal-Boundary Intersection, is used here to find evenly-spaced points on the Pareto curve for the `optimality' and `robustness' criteria. This Pareto curve enables the user to make the trade-off decision explicitly free of arbitrary `weighting' parameters. This paper also formulates a derivative-based approximation for evaluating the expected value of the objective function on the nonlinear manifold defined by the state equations for the system. Existing procedures for evaluating the expectation usually involve numerical integration techniques requiring many solutions of the state equations for one evaluation of the expectation. The procedure presented here bypasses the need for multiple solutions of the state equations and hence provides a cheaper and more easily optimizable approximation to the expectation. Finally, this paper discusses how nonlinear inequality constraints should be treated in the presence of random parameters in the design. Computational results are presented for finding a robust optimum of a structural optimization problem