Reinforcement Learning for Optimal Control of Network Epidemic Processes
Abstract
Our society is increasingly interconnected, making it easy for cascades/epidemic (diseases, disinformation etc). Current epidemic control efforts are based on approximate network epidemic models, which often ignore the unique complexity and rich information embedded in the complex interconnections of real-world networks/populations.Deep reinforcement learning (RL) is a powerful tool at learning policies for these nonlinear, complex processes in high-dimension. To control an epidemic outbreak on a Susceptible-Infected-Susceptible network epidemic model, we design a RL framework with a custom reward structure using the \textit{node2vec} embedding technique. Results indicate deep RL is able to determine and converge on an optimal intervention policy in a relatively short time.