Learning to Communicate with Limited Co-design

In this work we examine the problem of learning to cooperate in the context of wireless communication. We consider the two agent setting where agents must learn modulation and demodulation schemes that enable them to communicate with each other in the presence of a power-constrained additive white Gaussian noise (AWGN) channel. We investigate whether learning is possible under different levels of information sharing between distributed agents that are not necessarily co-designed. We make use of the “Echo” protocol, a learning protocol where an agent hears, understands, and repeats (echoes) back the message received from another agent. Each agent uses what it sends and receives to train itself to communicate. To capture the idea of cooperation between agents that are “not necessarily co-designed,” we use two different populations of function approximators – neural networks and polynomials. In addition to diverse learning agents, we include non-learning agents that use fixed standardized modulation protocols such as QPSK and 16QAM. This is used to verify that the Echo approach to learning to communicate works independent of the inner workings of the agents, and that learning agents can not only learn to match the communication expectations of others, but can also collaboratively invent a successful communication approach from independent random initializations. In addition to simulation-based experiments, we implement the Echo protocol in physical software-defined radio experiments to verify that it can work with real radios. To explore the continuum between tight co-design of learning agents and independently designed agents, we study how learning is impacted by different levels of information sharing – including sharing training symbols, sharing intermediate loss information, and sharing full gradient information. The resulting learning techniques span supervised learning and reinforcement learning. We find that in general, co-design (increased information sharing) accelerates learning and that this effect becomes more pronounced as the communication task becomes harder.