Speaker
Description
Personalized non-invasive neurostimulation is a candidate tool for closed-loop seizure control. We ask whether a reinforcement learning (RL) agent can learn when and how strongly to stimulate to stop seizure-like activity in a model network. We wrote a network of 100 coupled Morris–Lecar neurons, extended with calcium dynamics, as a TorchRL environment. Each agent step in the environment advances the differential equations with torchdiffeq. The agent sees only a time-delay embedding of population-level measures, as an electrode would. The reward weighs seizure-free time against stimulation strength and charge balance. A PPO agent learned state-dependent multi-pulse stimulation that ended collective oscillations in a growing share of the 200 parallel network instances. We also discuss the pitfalls: reward weights and batch sizes needed careful tuning, and the GPU speed-up came almost entirely from running many copies of the network side by side. A single network simulates faster on CPU, because the time steps run one after another and each is small.