[D] Benchmarking Deep RL Stability Capable of Running on Edge Devices

A new 'stable stack' for streaming deep reinforcement learning can process 477,000 observations per second on edge devices, using just 271k FLOPs per update. Tested on 433,000 real SSH attack logs, it uses JAX compilation and Exponential Moving Average normalization to handle sudden traffic bursts. The method relies on global gradient bounding for stability, proving critical for real-time, single-sample updates on non-stationary data streams where other techniques fail.