Spiking Neural Networks for Real-Time Strategy: A Curriculum-Trained SNN Agent for MicroRTS

Abstract

Spiking neural networks (SNNs) have demonstrated competence in board games, but their application to real-time strategy (RTS) games—which demand simultaneous multi-unit control, resource management, and long-horizon planning—remains unexplored. We present the first SNN agent capable of playing a complete RTS game. Our agent uses 2,494 leaky integrate-and-fire (LIF) neurons organized in a hierarchical architecture with a recurrent strategic core, trained via proximal policy optimization (PPO) with a three-stage curriculum. In a controlled comparison with parameter-matched multilayer perceptron (MLP) and gated recurrent unit (GRU) baselines across 5 random seeds, the unmodified SNN achieves the highest peak tournament performance but collapses in all 5 seeds during extended training, while a GRU baseline achieves perfect stability (\(100\%\) win rate, 0/5 collapses). A stabilized variant combining cosine learning rate decay, an entropy floor, and checkpoint reversion eliminates collapse entirely (0/5 seeds) and dominates a 12-agent tournament (\(89.5 \pm 8.8\%\) win rate, rank 1). A per-mechanism ablation identifies checkpoint reversion as the critical stabilizer. A timestep ablation shows that 8 internal steps match 16-step peak performance with reduced collapse (2/5 vs. 5/5), ruling out a pure compute-volume confound. The entire SNN agent fits within a single Loihi 2 neuromorphic core (2,494 of 8,192 neurons per core), avoiding the inter-core routing and synchronization overhead that is often the real bottleneck on neuromorphic chips, and suggesting that neuromorphic hardware can host competitive, stable RTS agents.