Stepwise uncertainty-aware offline reinforcement learning under constraints

Abstract

A computer-implemented method is provided for offline reinforcement learning with a dataset. The method includes training a neural network which inputs a state-action pair and outputs a respective Q function for each of a reward and one or more safety constraints, respectively. The neural network has a linear output layer and remaining non-linear layers being represented by a feature mapping function. The training includes obtaining the feature mapping function by constructing Q-functions based on the dataset according to an offline reinforcement algorithm. The training further includes tuning, using the feature mapping function, a weight between the reward and the one or more safety constraints, wherein during the obtaining and the tuning steps, an estimate of a Q-function is provided by subtracting an uncertainty from an expected value of the Q-function. The uncertainty is a function to map the state-action pair to an error size.