Universal attention-based reinforcement learning model for control systems

Abstract

A computing system trains a reinforcement learning model comprising multiple different attention model components. The reinforcement learning model trains on training data of a first environment (e.g., a first traffic intersection). The reinforcement learning model trains by training a state attention computer model on the training data that weighs each of respective inputs of a respective state. The reinforcement learning model trains by training an action attention computer model that determines a probability of switching from a first action to a second action of the first set of the multiple candidate actions (e.g., changing traffic colors of traffic lights). Alternatively, or additionally, a computing system generates an indication of a selected outcome according to the reinforcement learning model and sends a selection output to the second environment (e.g., a second traffic intersection with more lanes than the first traffic intersection) to implement the selected action in the second environment.