Quantum Reinforcement Learning for Dynamic Spectrum Access in Cognitive Radio Networks

Abstract
This thesis proposes Quantum Reinforcement Learning (QRL) as an improvement to conventional reinforcement learning-based dynamic spectrum access used within cognitive radio networks. The aim is to overcome the slow convergence problem associated with exploration within reinforcement learning schemes.
A literature review for the background of the carried out research work is illustrated. Review of research works on learning-based assignment techniques as well as quantum search techniques is provided. Modelling of three traditional dynamic channel assignment techniques is illustrated and the advantage characteristic of each technique is discussed. These techniques have been simulated to provide a comparison with learning based techniques, including QRL. Reinforcement learning techniques are used as a direct comparison with the Quantum Reinforcement Learning approaches. The elements of Quantum computation are then presented as an introduction to quantum search techniques. The Grover search algorithm is introduced. The algorithm is discussed from a theoretical perspective. The Grover algorithm is then used for the first time as a spectrum allocation scheme and compared to conventional schemes. Quantum Reinforcement Learning (QRL) is introduced as a natural evolution of the quantum search. The Grover search algorithm is combined as a decision making mechanism with conventional Reinforcement Learning (RL) algorithms resulting in a more efficient learning engine. Simulation results are provided and discussed. The convergence speed has been significantly increased. The beneficial effects of Quantum Reinforcement Learning (QRL) become more pronounced as the traffic load increases. The thesis shows that both system performance and capacity can be improved. Depending on the traffic load, the system capacity has improved by 9-84% from a number of users supported perspective. It also demonstrated file delay reduction for up to an average of 26% and 2.8% throughput improvement.