Searchability of Gene-Encoded Spiking Controllers under a Forward-and-Reproduction Energy Ledger

Training spiking neural networks (SNNs) is important for energy-efficient and biologically plausible computation. However, existing BPTT-based methods suffer from hard truncation, which breaks the eligibility trace and renders gradients unusable for propagation. Meanwhile, STDP-based methods avoid this gradient issue but provide no explicit task-driven direction, leaving the search trajectory undefined. To address this issue, we propose MED-SNN, an energy-constrained evolutionary framework that uses energy itself as the constraint driving search. In this way, energy shapes the search direction, while a gradient-free genetic algorithm performs optimization, thereby avoiding the gradient failure caused by hard truncation. The framework is gradient-free and directly encodes topology, weights, and metabolic parameters. It employs hard energy truncation—halting rollouts at the death threshold so that overspending trajectories leave the admissible set—and introduces inter-generational budget matching, which starts loose and progressively tightens toward the target energy ledger to preserve early selection signals. On CartPole-v1, MED-SNN remains searchable under training-time hard death (death rate 0.77). Hard-off: 409.9 ± 108.9 steps, 45% solve. Hard-on survival 0%. On the schedule table, matched-flat records hard-off 441.3 steps, 60% solve, and 100% hard-on survival; staged, linear, and cosine solve at 45%, 25%, and 45%. We claim training-time searchability under hard death.