Metaheuristic Optimization for Neuromorphic Computing Architectures

Abstract

Metaheuristic optimization has become an indispensable tool for neuromorphic computing architecture optimization, where the underlying decision problem - configuring spiking neural network topologies, synaptic parameters, and hardware mappings to minimize energy and latency on event-driven neuromorphic substrates - exhibits multimodality, high dimensionality, and constraint coupling that defy classical exact methods. This paper reviews and synthesizes more than two decades of progress on metaheuristic approaches to neuromorphic computing architecture optimization, covering the period 2000-2026. We organize the literature along four algorithmic families - physics-based, swarm-based, evolution-based, and hybrid - and discuss the principal operators and empirical signatures that distinguish each family. Particular attention is paid to recently proposed methods such as Evolution Strategies and Particle Swarm Optimization, whose reported gains on standard benchmarks merit careful reanalysis. Our synthesis draws on results from 5 benchmark families and applies a uniform protocol comprising pairwise Wilcoxon signed-rank tests and Friedman ranking with post-hoc critical-difference analysis. The principal findings are threefold. First, algorithmic progress has been real but incremental, with the largest gains attributable to operator-level innovations. Second, persistent challenges related to hardware-aware mapping and noise tolerance continue to motivate adaptive and hybrid schemes. Third, methodological variability across studies significantly limits cross-study comparability and calls for standardized benchmarks. We close with a roadmap identifying five interlocking challenges likely to shape the next decade, including surrogate-model integration, adaptive operator selection, and reproducibility safeguards. The reviewed evidence suggests that progress on these fronts would yield tangible benefits in energy per inference and spike count.