In this paper, we present a novel technique for reinforcement learning based spectrum sharing schemes. It shows how the potential action space of users is reduced by randomly partitioning the spectrum for different users. Cognitive agents are therefore able to finish their exploration stage faster than more basic reinforcement learning based schemes. Simulation results show how the communication system is more stable if all users reserve a subset of resources at the beginning and then learn to

more »

... elect the best channels in the reserved spectrum set. The system performance by using pre-partitioning is investigated and comparison with the traditional reinforcement learning based schemes is given.