Variable speed limit (VSL) control is a flexible way to improve traffic conditions, increase safety, and reduce emissions. There is an emerging trend of using reinforcement learning methods for VSL control. Currently, deep learning is enabling reinforcement learning to develop autonomous control agents for problems that were previously intractable. In this paper, a more effective deep reinforcement learning (DRL) model is developed for differential variable speed limit (DVSL) control, in which dynamic and distinct speed limits among lanes can be imposed. The proposed DRL model uses a novel actor-critic architecture to learn a large number of discrete speed limits in a continuous action space. Different reward signals, such as total travel time, bottleneck speed, emergency braking, and vehicular emissions are used to train the DVSL controller, and a comparison between these reward signals is conducted. The proposed DRL-based DVSL controllers are tested on a freeway with a simulated recurrent bottleneck. The simulation results show that the DRL based DVSL control strategy is able to improve the safety, efficiency and environment-friendliness of the freeway. In order to verify whether the controller generalizes to real world implementation, we also evaluate the generalization of the controllers on environments with different driving behavior attributes. and the robustness of the DRL agent is observed from the results.

变速限制（VSL）控制是改善交通状况，提高安全性和减少排放的灵活方法。使用强化学习方法进行VSL控制的趋势正在兴起。当前，深度学习使强化学习能够为以前难以解决的问题开发自主控制代理。本文针对微分变速限速（DVSL）控制系统开发了一种更为有效的深度强化学习（DRL）模型，该模型可以在车道之间施加动态且明显的限速。提出的DRL模型使用新颖的行为者批评体系结构来学习连续动作空间中的大量离散速度限制。诸如总行驶时间，瓶颈速度，紧急制动和车辆排放等不同的奖励信号用于训练DVSL控制器，并在这些奖励信号之间进行比较。提议的基于DRL的DVSL控制器在高速公路上经过模拟的经常性瓶颈测试。仿真结果表明，基于DRL的DVSL控制策略能够提高高速公路的安全性，效率和环境友好性。为了验证控制器是否可以推广到现实世界中，我们还评估了具有不同驾驶行为属性的环境下控制器的推广。从结果可以看出DRL剂的坚固性。高速公路的效率和环境友好性。为了验证控制器是否可以推广到现实世界中，我们还评估了具有不同驾驶行为属性的环境下控制器的推广。从结果可以看出DRL剂的坚固性。高速公路的效率和环境友好性。为了验证控制器是否可以推广到现实世界中，我们还评估了具有不同驾驶行为属性的环境下控制器的推广。从结果可以看出DRL剂的坚固性。