Learning Power Control From a Fixed Batch of Data

Abstract

We address how to exploit historical power control data, gathered from a monitored environment, for accelerating the learning of power control in an unexplored environment when only partial channel state information, e.g., path-loss, is available. We adopt offline deep reinforcement learning (DRL), whereby the agent learns the policy to produce the transmission powers by using the historical data and occasional exploration and develops a new algorithm called modified batched constrained Q-learning (mBCQ). Compared to conventional continuous DRL algorithms, mBCQ increases the learning speed by almost 50 times and demonstrate robustness to hyper-parameters.