DYNAMIC SPECTRUM ACCESS WITH NON-STATIONARY MULTI-ARMED BANDIT

Dynamic spectrum access (DSA) is an emerging notion in cognitive radio, aiming to improve the spectrum usage with reliable secondary access to the spectral resources. The main challenge in DSA is the detection of spectral opportunities and their efficient utilization without causing interference to the primary users. For this goal, we propose to make use of a reinforcement learning approach: the Multi Armed Bandit (MAB). The MAB approach provides the secondary users with the rules and policies necessary to achieve a tradeoff between exploitation and exploration in DSA. Different MAB strategies are tested on an IEEE802.11 medium access model and evaluated in dynamic environment. Our study shows that the MAB constitute a viable solution for the DSA. Adding to that, the performances of the MAB algorithms can be improved with a finite tuning of the internal parameters.