Title:Risk averse constrained blackbox optimization under mixed aleatory/epistemic uncertainties

Abstract:This paper addresses risk averse constrained optimization problems where the objective and constraint functions can only be computed by a blackbox subject to unknown uncertainties. To handle mixed aleatory/epistemic uncertainties, the problem is transformed into a conditional value-at-risk (CVaR) constrained optimization problem. General inequality constraints are managed through Lagrangian relaxation. A convolution between a truncated Gaussian density and the Lagrangian function is used to smooth the problem. A gradient estimator of the smooth Lagrangian function is derived, possessing attractive properties: it estimates the gradient with only two outputs of the blackbox, regardless of dimension, and evaluates the blackbox only within the bound constraints. This gradient estimator is then utilized in a multi-timescale stochastic approximation algorithm to solve the smooth problem. Under mild assumptions, this algorithm almost surely converges to a feasible point of the CVaR-constrained problem whose objective function value is arbitrarily close to that of a local solution. Finally, numerical experiments are conducted to serve three purposes. Firstly, they provide insights on how to set the hyperparameter values of the algorithm. Secondly, they demonstrate the effectiveness of the algorithm when a truncated Gaussian gradient estimator is used. Lastly, they show its ability to handle mixed aleatory/epistemic uncertainties in practical applications.