Multi-agent Q-learning for autonomous D2D communication

This paper is devoted to autonomous device-to-device (D2D) communication in cellular networks. The aim of each D2D pair is to maximize its throughput subject to the minimum signal-to-interference-plus-noise ratio (SINR) constraints. This problem is represented by a stochastic non-cooperative game where the players (D2D pairs) have no prior information on the availability and quality of selected channels. Therefore, each player in this game becomes a “learner” which explores all of its possible strategies based on the locally-observed throughput and state (defined by the channel quality). Consequently, we propose a multi-agent Q-learning algorithm based on the players' “beliefs” about the strategies of their counterparts and show its implementation in a Long Term Evolution - Advanced (LTE-A) network. As follows from simulations, the algorithm achieves a near-optimal performance after a small number of iterations.