Research On Off-line Learning Technology For Cognitive Radio Network

Cognitive radio (CR) is a kind of intelligent radio communication system,which can provide a feasible idea to resolve the contradiction between wirelessbusiness requirements and spectrum resource shortage. As the core of CRtechnology, the intelligent learning and reasoning plays decisive role foreffectively using the spectrum resources, meeting user needs and improvingsystem performance. Cognitive engine (CE) can be designed by introducingreasoning and learning methods from the field of artificial intelligence, whichaims at realizing parameters re-configuration. This paper focuses on studyingoff-line learning technology for cognitive radio network, which maily includeslearning, reasoning and deciding for cognitive engine. The main works aresummarized as follows.Firstly, the paper proposes a scheme of CR optimal engine based on DiscreteUniform Genetic Algorithm (DUGA). As the parameter adjustment for CR is atypical multi-objective optimization problem, the paper presents the method,considering the optimization of jointing PHY layer and MAC layer. Comparedwith traditional genetic algorithm, the DUGA applies the principle of discreteuniform distribution to effectively overcome the problems of falling into thelimitations of local optimization in the wheel plate selection. Crossover,mutation and gradual iteration are employed to achieve population diversity andrapid convergence. The MATLAB simulation results show that the performanceof the DUGA is better than the typical NNIA. The established optimizationengine model is applied to the CR communication system under the NS2. Theresults show that the optimal engine can effectively enhance the systemthroughput and reduce average transmission power.Secondly, in this paper, a multi-layer network learning model is proposedbased on RBF neural network, which are effectively applied to the design ofcognitive engine. The learning model is different from the traditional neuralnetwork learning model, which is divided into outer and inner network to isolatethe learning of global and local parameters. The information flow are drove fromthe outer layer to the inner layer, which can improve the forecasting ability, reduce the learning time and increase the adaptive capacity to the environmentfor the network. The perception data are collected from NS2platform to simulatewireless communication systems, which are employed to train network. Theperception information is employed to train and test the outer and inner network.The outer layer neural network learns global parameters to reconfigure therouting protocol, while the inner layer neural network configures localcommunication parameters according to the routing protocol. The multi-layerlearning model is applied to reconfigure CR communication system based onperception information and user needs. Simulation results show that the learningmodel can effectively adapt to environmental change and meet user needs.Finally, two cognitive engine algorithm modes are established based on asummary of the research at earlier stage, which are RS-RBF and GA-RBFlearning model, separately. Learning models combines rough sets and geneticalgorithm with RBF neural network, which are applied to CR cognitive enginedesign. Rough sets can effectively deal with the redundancy, uncertainty andincompleteness of communication system, which will reduce data storage spaceto improve the speed of network training. Operating parameters are selectedempirically in the learning process of RBF neural network, which may lead toaffect the network adaptability. The perception data is collected from NS2, whichis employed to train learning model. Simulation results show that the learningmodel can effectively achieve CR learning reconfiguration, meet user needs andimprove the network adaptability.