Reinforcement learning based decision-making model in autonomous vehicle control for cooperation and mitigation of collision among multiple vehicles

• Main Author: Abu Jafar, Md Muzahid
• Format: Thesis
• Language: English
• Published: 2022
• Subjects: Q Science (General); QA75 Electronic computers. Computer science
• Online Access: http://umpir.ump.edu.my/id/eprint/37642/1/ir.Reinforcement%20learning%20based%20decision-making%20model%20in%20autonomous%20vehicle%20control%20for%20cooperation%20and%20mitigation%20of%20collision%20among%20multiple%20vehicles.pdf; http://umpir.ump.edu.my/id/eprint/37642/

Self-driving cars have become a popular research topic in recent years. Autonomous driving is a complicated field of study that involves a variety of disciplines, such as electronics, computer vision, geo-location, decision-making, or control. Autonomous vehicles are an example of non-linear technologies being used in the real world. Controlling this kind of device in particular situations in the context of multi-agent traffic systems is difficult because of instability. This type of equipment demands expertise, and it is even more difficult to create this understanding of talent as an independent control system. Because each agent has its own self-determined protocol decision management, it is hard to coordinate several autonomous devices on a single job. Over the last decade, there has been a lot of attention on sequential decision-making under ambiguity and uncertainty, which is a distinct range of challenges requiring an agent to interact with an uncertain environment to achieve a target. Reinforcement learning methods applied to these challenges have resulted in recent AI achievements in robotics, game playing, and other areas. In response to these empirical testimonies, this project confronts the problem of multiple vehicle control decisions and performs control strategies for the avoidance of severe multiple vehicle collisions in autonomous vehicles. These control techniques rely on the reinforcement learning model and deploy two distinct traffic scenarios for progressing research flow. An extensive taxonomy conveyed the existing protocols and solutions, and a conceptual model for MVCCA was formulated first. Then, using the Reinforcement Learning-based Decision- Making (RLDM) model, the system is developed and implemented. An extensive simulation gives us the best outcomes for the development of optimum driving strategies in a multi-agent traffic environment. We extensively evaluate the training performance, driving performance, and the ability of collision avoidance as well. We investigated the training performance of both the single vehicle and multiple vehicle environments. Validation of the decision-making scheme would create new opportunities for autonomous driving, as well as new concepts and applications.