Yazılım tanımlı bir ağ mimarisinde derin takviyeli öğrenmeyikullanarak kablosuz sensör ağlarının ömür boyu optimizasyonunu sağlama

According to the changing topologies of modern networks, the use of static routing rules has become obsolete. Software-Defined Networks (SDNs) are being used to overcome such limitation, where a central controller handles the decision-making role regarding packets routing. This controller collects information about the network, in addition to the packet information, to decide the route a packet should follow to reach its destination. However, with the growing complexity of WSNs topologies and the importance of efficient routing, Machine Learning (ML) techniques are being used to handle the decision making in the SDN controller. In this study, a new method is proposed to optimize the resources consumption in a WSN that uses SDN. The proposed method employs a neural network that is trained using Reinforcement Learning (RL), based on the lifetime of the WSN. To extend the lifetime of WSN the neural network is required to optimize the power consumption of the nodes in that network, in which the optimal routes must be used. Three types of neural networks are evaluated in this thesis; Feed-Forward Neural Network (FF-NN), 2D-Convolutional Neural Network (2D-CNN) and 3D-CNN. The evaluation of these models show that the using the 3D-CNN has achieved the best performance, with an average lifetime of 678251.6 seconds, with an extension of 17% of the 578122.2 seconds using the existing state-of-the-art method. The average number of hops a packet is required to travel through, to reach its destination, in this model is 9.81 hops with an average Packet Delivery Rate (PDR) of 85.07%. Additionally, the 2D-CNN model has achieved 638169.2 seconds lifetime, with an average of 12.37 hops per packet and 82.47% PDR, whereas the FF-NN has achieved 578381.6 seconds lifetime with 83.37% PDR and 8.31 hops per packet. In addition to the superiority of the 3D-CNN, the results also show that the use of the shortest paths causes an exhaustion to the resources of certain nodes, positioned in locations that handle extensive traffic, which reduces the overall lifetime of the WSN. Thus, the extension of the lifetime requires using alternative, i.e. longer, paths to avoid such exhaustion and extend the lifetime of the network.