ايمن بشير يوسف
The Internet is a shared resource in which users compete for limited network capacity. Contention between separate user requests can cause congestion, which might result in significant queue delays, packet losses, or both. Congestion control limits the pace at which traffic sources inject packets into a network in order to maximize bandwidth consumption while minimizing network congestion. Currently, there are two techniques to handle congestion, the end-toend mechanism which is achieved by the Transition Control Protocol (TCP) and middle point algorithms as an AQM in routers. AQM algorithm is one of the most significant study areas in network congestion control; nevertheless, new selflearning network management algorithms are needed on nodes to cope with the huge quantity of traffic and minimize queuing latency. For efficient network management, AQM has been used in machine learning to auto adapts the parameters of algorithms on Deep Reinforcement Learning. This work study an AQM based to the deep reinforcement learning on handle arrival time and the choose between production and waiting. In the current research, the Deep Q management is selected for the linear model of (TCP/AQM) system that is represented by Python program. The system is based on an intelligent agent that learns by adjusting behaviors and measuring results. Over time, the system study by picking the optimal actions. Also, Network simulation 3 is used to support the design of controller. It has been found that the system performs better when compared to AQM schemes (RED) algorithm, the results obtained through 10 different experiments have proven that the DQN algorithm is able to maintain the queue length close to the desired value by 92%, reduce the drop rate by 7%, and increase the throughput by 0.2Mbps under different network traffic conditions (load factor, bottleneck link and data rate) compared with RED algorithm. Comparing the results with the previous works [15], it was noticed that the results of this study algorithm are better in terms of throughput, loud factor, and drop II rate. Where the double of the client in the previous study was used, and this is because of the number of times the algorithm was trained, which enabled it to obtain better results.As well as in terms of the throughput, where increasing the layers of perceptron will study the performance of the algorithm in terms of the throughput by an increase of 0.17Mbps, as well as 5% better in terms of the drop rate.
