Toward Order Optimal Channel Access in Unknown Environments: An Online Learning Method

We investigate the opportunistic channel access problem in a centralized cognitive radio network, where the cognitive base station (CBS) periodically detects spectrum holes in the licensed network and coordinates the unlicensed secondary users to utilize the idle channels. Existing spectrum access mechanisms typically rest on the assumption that the spectrum environment is known in advance, i.e., licensed channel states exhibit either stochastic or adversarial variations across different time instances and locations. These approaches address the spectrum access problem from the perspective of online learning, which relies heavily on the prior knowledge of network parameters such as the time horizon and the user activities to tune hyperparameters. In this paper, we tackle the multiuser channel access task by formulating it as a combinatorial multi-armed bandit problem in an unknown environment, where we propose an online mirror descent-based channel access method for the CBS that does not require any prior knowledge of the environment. Our proposed method adaptively adjusts the probability of secondary users accessing the licensed channels based on the historical transmission feedback, achieving order-optimal performance in both stochastic and adversarial environments. Numerical results validate the theoretical analysis and also demonstrate that the proposed method outperforms others under various network settings.