Optimisation problems often involve uncertain parameters whose distribution shifts in response to the decision itself. When only a stochastic simulator is available and the number of draws is fixed, minimising a spectral risk objective—expected cost plus a weighted conditional value-at-risk—becomes challenging. The paper proposes adaptive conditional residual sampling (ACRS) to address this setting.
ACRS begins by taking one simulator draw at each point of a space-filling design, then fits a heteroscedastic regression model of the decision-dependent distribution. Standardised residual vectors are reused as scenarios to build a deterministic surrogate. The algorithm searches this surrogate for several local optima, adds new draws around the most promising candidates, ranks them on the simulator using common random numbers, and finishes with sample-path local refinement via tail importance sampling.
Theoretically, the optimality gap of the surrogate minimiser is bounded by the Wasserstein error of the scenario distribution, and the refinement estimator is consistent. These results provide formal support for the heuristic components of the method.
In experiments with a shared budget accounting, design sampler, and refinement routine, ACRS achieved a mean regret of 11.8% on 60 random analytic instances, compared with 22.3–42.3% for Bayesian optimisation, cross-entropy, kernel-density, stochastic-gradient, and refinement-only baselines. Its advantage was largest at small budgets and tended to grow with dimension relative to the kernel-density method. ACRS also had the lowest mean regret on two six-dimensional benchmarks and three development problems, while performing comparably to pure local refinement on three held-out problems.