Methods and systems for training an untrained policy network of an autonomous agent to generate a dual-action policy network

Abstract

A computer-implemented method for training untrained policy network 204 of an autonomous agent using a simulator to obtain a dual-action policy network. The method comprises: a) Randomly generating an initial state of training environment 100. b) With the policy network, determining a simulated action and a real action to be taken by the autonomous agent base on the initial state. c) Updating the environment based on the simulated action to generate a next state. d) Determining an intrinsic reward representative of a degree of similarity between the simulator and the real world. e) Determining a modified task reward based on an original task reward and the intrinsic reward. f) Updating the policy network based on the modified task reward. And g) repeating steps b) to f) based on the updated environment until a predetermined condition is met, wherein the next state is used instead of the initial state.

Images