Abstract: Nature-inspired metaheuristic algorithms have become increasingly popular in the last couple of decades, and now constitute a major toolbox for tackling complex high-dimensional optimization problems. Using group sequential experimentation, adaptive design, multi-armed bandits, and bootstrap resampling methods, this study develops a novel statistical methodology for efficient and systematic group sequential selection of the tuning parameters, which are widely recognized as pivotal to the success of metaheuristic optimization algorithms in practice, as new information accumulates during the course of an experiment. The methodology is applied to compute optimal experimental designs in nonlinear regression models, and is illustrated with solutions of long-standing optimal design problems in early-phase dose-finding oncology trials.

Key words and phrases: Adaptive group sequential designs, compound optimality criterion for toxicity and efficacy, locally D-optimal and c-optimal designs.