Complexity of Spectrum Activity and Benefits of Reinforcement Learning for Dynamic Channel Selection

Abstract

We explore the question of when learning improves the performance of opportunistic dynamic channel selection by characterizing the primary user (PU) activity using the concept of Lempel-Ziv complexity. We evaluate the effectiveness of a reinforcement learning algorithm by testing it with real spectrum occupancy data collected in the GSM, ISM, and DECT bands. Our results show that learning performance is highly correlated with the level of PU activity and the amount of structure in the use of spectrum. For low levels of PU activity and/or high complexity in its utilization of channels, reinforcement learning performs no better than simple random channel selection. We suggest that Lempel-Ziv complexity might be one of the features considered by a cognitive radio when deciding which channels to opportunistically explore.