Reinforcement Learning for Mixed Cooperative/Competitive Dynamic Spectrum Access

A dynamic spectrum sharing problem with a mixed collaborative/competitive objective and partial information about peers’ performances that arises from the DARPA Spectrum Collaboration Challenge is considered. Because of the very high complexity of the problem and the enormous size of the state space, it is broken down into the subproblems of channel selection, flow admission control, and transmission schedule assignment. The channel selection problem is the focus of this paper. A reinforcement learning algorithm based on a reduced state is developed to select channels, and a neural network is used as a function approximator to fill in missing values in the resulting input-action matrix. The performance is compared with that obtained by a hand-tuned expert system.