Adaptive Evolutionary Search in Changing Environments: Moving Peaks and a Two-Species Chemostat Surrogate

Abstract

Dynamic optimization can support the search for operating designs when a simulated nutrient regime changes during an optimization run. This study reports AMPERE, a memory-based evolutionary search combining sentinel monitoring, a regime-indexed archive, empirical transition estimates, allocation of search effort and a magnitude-scaled mutation response. The reported campaign compares six methods on a moving-peaks landscape with 62,500 evaluations and a fixed two-species chemostat surrogate with 17,500 evaluations, using ten main-comparison seeds and five ablation seeds. On moving peaks, AMPERE has a larger within-epoch incumbent-drop metric than the plain GA (0.293 ± 0.124 versus 0.103 ± 0.031; Holm p = 0.0098), but a smaller value than fixed triggered hypermutation (2.622 ± 0.788). On the chemostat surrogate, its mean current-best endpoint titer is 4.08 ± 0.17 g/L, compared with 1.55 ± 0.41 g/L for the plain GA. Its comparisons with random immigrants and triggered hypermutation do not reject equal performance at the reported significance level. Removing AMPERE’s mutation response gives 1.76 ± 0.50 g/L in the reported ablation; the five-pair two-sided p-value is 0.0625. These observations show benchmark-dependent performance without establishing an isolated benefit of anticipation or memory. The incumbent-drop metric is not regret to a known optimum, and neither benchmark validates biological adaptation or online plant control. Exact simulator parameters, paired outputs, event counters and implementation details remain necessary for independent reproduction.