Deep Reinforcement Learning for Contagion Control

In this work, we present a networked epidemic model comprising non-identical agents and consider the problem of learning an allocation strategy to contain an outbreak. Even though spreading processes are generally described by nonlinear dynamics, most methods for control are typically based on linear approximations of the nonlinear process and assume full knowledge of the propagation model and dynamics. We propose an alternative approach based on deep reinforcement learning. We define an environment to represent a heterogeneous nonlinear model and show that this environment can be used in conjunction with a Deep Q-Network to stabilize the spreading process. We illustrate our approach using real data from an air traffic network.