Distributed Algorithm for Learning to Coordinate in Infrastructure-Less Network

We consider the spectrum access in an unlicensed spectrum (i.e., no incumbent users) for infrastructure-less networks where the number of users are unknown and they cannot coordinate with others due to lack of a control channel or a central controller. Also, users do not have spectrum sensing capability due to size and power constraints in battery operated radios. Such a setup is being studied for Internet of Things applications to enable sensors to communicate sensed data without the need of dedicated spectrum and network infrastructure. Using multi-user multi-armed bandit-based learning framework, we propose a new distributed algorithm which achieves a lower regret (i.e., throughput loss) than existing algorithms while keeping the number of collisions low. Fewer collisions save power which would have been otherwise wasted due to re-transmissions. High confidence bounds on the regret and number of collisions along with simulation results validate the effectiveness of our algorithm.