Cognitive Computing for Smart Spectrum Allocation in Wireless Systems

To fill the gap of dynamic and efficient spectrum management, this research presents a cognitive computing for smart spectrum allocation in wireless systems. In preprocessing, the input features are uniform scaled such that the input features are scaled using Z-Score Standardization for learning stability. Diminsionality reduction is carried out with Principal Component Analysis (PCA) to use less data, thus reducing the computational complexity while maintaining the important data patterns. A Multi Agent Deep Q Network (MADQN) makes up the core decision making component and allows the distributed agents to learn optimum spectrum allocation strategies in a decentralized manner. A model is built using TensorFlow for training, and the wireless communication environments and reinforcement learning interactions are simulated using NS-3 and OpenAI Gym. Results of simulation indicate that spectrum utilization and collision avoidance and convergence time increase from traditional methods are quite large. Finally, a fully integrated approach to spectrum decision making is presented; it extends from raw channel measurements, through autonomous inference discovery, to intelligent adaption of network configuration based at MAC, link and network levels using offline and continuous online methods.