Distributed Algorithms for Efficient Learning and Coordination in Ad Hoc Networks

A distributed sampling strategy for multiple (N) agents is considered that minimizes the sample complexity and regret of acquiring the best subset of size $N$ among total $K$ ≥ $N$ channels in a cognitive radio access setup. Agents cannot directly communicate with each other, and no central coordination is possible. Each agent can transmit on one channel at a time, and if multiple agents transmit on the same channel at the same time, a collision occurs, and no agent gets any information about the channel gain or how many other agents transmitted on the same channel. If no collision occurs, the agent observes a reward (or gain) sample drawn from an underlying distribution associated with the channel. An algorithm to minimize the sample complexity and regret is proposed. One important property of our algorithm that distinguishes it from the prior work (that do not assume knowledge of N) is that it requires no information about the difference of the means of the channel gains of the $K$ channels. Our approach results in fewer collisions with improved regret performance compared to the state-of-the-art algorithms. We validate our theoretical guarantees with experiments.