Redundancy-Triggered Relocation: Decentralized Coordination of Local Searchers for Static and Dynamic Multimodal Optimization

Abstract Dynamic multimodal optimization asks for many optima, again after every change, and teams of local searchers waste budget when several converge to the same basin. In redundancy-triggered relocation (RTR), a searcher that perceives a fitter teammate within an exclusion radius restarts elsewhere. Each searcher perceives only the teammates within its own radius, so redundancy is judged at a small scale without a radius tuned to the problem, and no clustering or central controller is needed. We show that global exclusion caps the number of optima a team can hold, and that limited perception lifts this cap. With covariance matrix adaptation evolution strategy (CMA-ES) searchers on the organisers' instances of the CEC 2022 dynamic multimodal benchmark, RTR raises the mean peak ratio from 0.503 to 0.628, or 0.640 without two optional components, and ranks first among the dynamic baselines we ran. It trails the best published entries of the 2022-2024 competitions (0.648 to 0.686), but has the highest mean on the 10-dimensional cone landscapes. Controls with criteria fixed before their runs point to the trigger as the source of the gain, since relocating at random times at the same rate, or restarting on convergence even with an archive, is worse on 20 to 23 of 24 problems, and teammate information delayed by up to one cycle does not measurably hurt. On the CEC 2013 niching benchmark it is close to the best published results on the 5- to 20-dimensional composition functions, but behind recent methods on low-dimensional problems with many optima.