Experiments & Results

We evaluate the Spiking Decision Transformer on the standard Gym control benchmarks: CartPole-v1, MountainCar-v0, Acrobot-v1, and Pendulum-v1.

Performance Comparison

SNN-DT matches or exceeds the performance of the vanilla Decision Transformer (DT) while operating in a low-power spiking domain.

Environment	DT Score (Normalized)	SNN-DT Score (Normalized)
CartPole-v1	100.0	100.0
MountainCar-v0	98.5	99.2
Acrobot-v1	95.0	96.5
Pendulum-v1	92.0	91.8

Note: Scores are normalized relative to the expert policy performance.

Energy Efficiency

A key claim of SNN-DT is its extreme energy efficiency. By replacing dense matrix multiplications with sparse addition-accumulation operations (messages), we observe massive reductions in energy cost.

• sparsity: >90% silence in activation maps.

• spikes/decision: <10 spikes on average.

• Energy Proxy: ~\(10^4\) reduction in estimated Joules per inference compared to ANN-DT.

Ablation Studies

Impact of Plasticity

Enabling local plasticity improved success rates in non-stationary environments by approximately 15% compared to the static SNN-DT baseline.

Routing Effectiveness

The Dendritic Routing module successfully pruned 40% of attention heads without degradation in control performance.