Cooperation in wireless networks: a game-theoretic framework with reinforcement learning

A game-theoretic framework based on the iterated prisoner's dilemma (IPD) is proposed to model the repeated dynamic interactions of multiple source nodes when communicating with multiple destinations in an ad hoc wireless network. In such networks where nodes are autonomous, selfish and not familiar with other nodes' strategies, fully cooperative behaviours cannot be assumed. Therefore reinforcement learning is studied to relate the utility function of each source node to actions previously taken in order to learn a strategy that maximises their expected future reward. Particularly, a Q-learning algorithm is proposed to allow network nodes to adapt to and play the IPD game against opponents with a variety of known and unknown strategies. Simulation results illustrate that the proposed Q-learning algorithm allows network nodes to play optimally and achieve their maximum expected return values.