Reinforcement learning systems for controlling wireless communication networks

Abstract

A computer implemented method of training a reinforcement learning model for controlling a dynamic system includes generating a trajectory sample of a simulated system that corresponds to the dynamic system, the trajectory sample including a current state sof the simulated system at time t, an action ataken on the simulated system at time t according to a policy π, a subsequent state sof the simulated system following the action a, and a reward r associated with the action at, and estimating a robust target value V(s) for the trajectory sample, wherein the robust target value V(s) includes an expected value of a sum of the reward r and a minimum estimated value V(s) of the simulated system at the subsequent state sbased on a plurality of transition possibilities p from the current state sin response to the action a. The method updates a value function estimator based on the robust target value, and updates the policy based on the trajectory and the value function estimator.