Route Optimization using Hybrid GRU Learning Model for SDN and Edge-Based VANET Topology

Journal of computer sciences

article

Savitha Kumar¹  Chandrasekar Chinnasamy² 

[1] Department of Computer Science, Government Arts College;Department of Computer Science, Government Arts and Science College for Women

关键词: SDN;    VANET;    ML;    Reinforcement Learning;    GRU;    Edge Computing;   

DOI  :  10.3844/jcssp.2022.743.756

学科分类：计算机科学（综合）

来源: Science Publications

【 摘 要 】

Intelligent Transportation System (ITS) offersoutstanding features, including security applications and emergency alerts.Unfortunately, ITS limits traffic control services, adaptability, andadjustability due to the traffic volume. Hence, expanding the standard VehicularAd hoc Networks (VANET) framework is a requirement. As a result, in the latestdays, the concept of Software-Defined Networking based Vehicular Ad hocNetworks (SDN-VANETs) has drawn considerable attention, creating VANETssmarter. The SDN-VANETs design is capable of addressing the aforementionedVANET issues. The integrated (analytically) SDN architecture is customizableand it also contains domain knowledge about the VANET architecture. Packetforwarding is a fundamental challenge in VANET wherein a router, as in the formof an RSU, determines the next hop of every signal in the pipeline to provideit to its recipient as fast as possible. Reinforcement Learning (RL) is used todevelop autonomous routing protocol rules; however, the limitation of RL'sdepth prevents it from representing more comprehensively dynamic networkconditions, restricting its true value. In this research, we present a VANETinfrastructure based on SDN + EDGE with a new Deep Reinforcement Learning (DRL)route optimization framework, "Gated Recurrent Reinforcement Learning(GRRL)" neural network, whereby each router has its hybrid GRU + FeedForward Neural Network (NN) for learning and making decisions in an entirelydistributed environment. The GRRL collects routing characteristics fromvaluable information about huge backlog packets and previous operations,substantially approximating the weighting scheme of Q-learning. We also enableevery route to connect with its immediate neighbors regularly such that a morecomprehensive view of network topology may be integrated. Trial findingsdemonstrated that the multi-agent GRRL strategy could achieve a delicatebalance between congestion awareness and the fastest routes, considerablyreducing packet transmission time in general network topologies compared to itscompetitors.

【 授权许可】

CC BY