A model-based deep learning approach to spectrum management in distributed cognitive radio networks

Abstract

The acceleration towards the fifth generation (5G) and beyond will see the internet of things (IoT) being the primary strategy of deployment, and wireless networks will become more distributed and autonomous. Furthermore, network users will demand delivery of multimedia content to various network devices in dissimilar contexts. Thus, the cognitive radio (CR) paradigm requires some improvements for it to rigorously resolve quality of service (QoS) and quality of experience (QoE) in an energy-efficient manner before the 5G network is commissioned. Therefore, solving the distributed RA problem through thorough and in-depth investigations into the essentials and intricacies of energy-efficient RA by integrating artificial intelligence (AI) and signal processing (SP) strategies is a requisite. Having identified this knowledge gap and several limiting factors, this thesis focuses on two fronts to maximize the distributed opportunistic usage of the wireless spectrum with enhanced energy efficiency. The first contribution of this study provides a solution for missing spectrum sensing information to improve spectrum occupancy measurements in distributed CRNs. This is a problem commonly encountered in distributed cooperative spectrum sensing scenarios, where secondary users (SUs) are faced with the missing spectrum sensing data (SSD) problem owing to several impairments such as (i) the use of specific collaborative spectrum sensing schemes and (ii) imperfect reporting channel conditions. This results in the SSD contributed by SUs having gaps of missing entries. This degrades the performance of spectrum sensing algorithms, especially when the amount of missing SSD is quite large. Therefore, spectrum occupancy reconstruction is proposed as a solution to deal with missing values through missing value imputation. This is a deep learning (DL)-based strategy that uses deep belief networks (DBNs) composed of restricted Boltzmann machines (RBMs) to capture the feature of the input space of the spectrum occupancy data from a Markov random field (MRF). Link energy functions from the Ising models and the Metropolis-Hastings algorithm are used to pre-train the RBM to obtain a spectrum occupancy data matrix. The size of training samples and learning rates are decided using Gibbs sampling during the training process and missing spectrum values are learned using a scaled stochastic gradient descent (SGD). The simulation results obtained indicate that spectrum occupancy reconstruction problems can be solved better using the SGD algorithm because it takes advantage of correlations in multiple dimensions better than singular value decomposition (SVD) in matrix factorization. The second contribution provides a solution for energy saving and QoS provisioning for SUs with heterogeneous traffic, which is a problem exacerbated by the increased demand for multimedia services. This necessitates for the establishment of newer power control strategies for multimedia sources, where energy saving and QoS provisioning are viewed from the job arrival rate instead of the packet arrival rate perspective. Here, the model dynamics are formulated as a continuous-time non-linear input affine system which combines opportunistic transmission and opportunistic computing to obtain resource consumption efficiency. By treating the base station (BS) as a hybrid switching system, a weighted cost function is obtained and solved using model-based reinforcement learning (RL), which initiates a single look-ahead for optimum operating states. Then, using the resource consumption efficiency, a DL-based predictive control scheme was realized with control actions that drive a stacked auto-encoder (SAE) that plays dynamic games on queues and performs effective trade-offs between QoS provisioning and energy saving. The simulation results obtained indicate that the processor sharing (PS) scheduling scheme achieves better energy saving than first-come-first-served (FCFS) at higher job arrival rates. The last contribution deals with the problem of distributed RA in energy-constrained CRN environments, with the objective of ensuring user satisfaction in terms of QoE and QoS in an energy-efficient manner. QoE evaluation is performed using docitive techniques and the results obtained indicate that transfer-learning through docitive approaches achieves better convergence rates and superior spectral efficiency compared to the traditional cognitive approaches. Then, a computationally efficient optimization technique that handles the energy efficiency learning model is achieved using factored Markov decision processes (FMDPs), which provides a solvable framework for energy minimization. This completes the hierarchical deep RL (DRL) with a deep Q-network (DQN) formulation that learns energy consumption subject to latency constraints. The results obtained show that the DQN approach with experience replay achieves better QoS performance compared to the traditional RL in terms of minimizing buffer delays and power consumption.