Reinforcement Learning over Noisy Channels: An Information Bottleneck Approach

In this work, we study reinforcement learning (RL) in a remote control wireless setting, in which a transmitter observes the state of an environment, processes and transmits a message to a receiver, which then takes an action in the environment based on the received message. We take a goal-oriented communications approach to encoding (and decoding) the transmitted message with the purpose of maximizing reinforcement learning performance. We apply the information bottleneck (IB) approach to training the encoder and decoder, which uses information-theoretic concepts to balance the compression of the input data with the learning performance. Using the familiar deep Q-network (DQN) RL approach, we provide an information theoretic interpretation of the IB approach and the noisy channel. Simulations demonstrate the training performance and generalization performance under various settings of the IB tradeoff parameter, the bottleneck dimension, and channel noise.