基于强化学习的自动驾驶联合训练方法

Joint Training Method for Autonomous Driving Based on Reinforcement Learning

自动驾驶控制算法对自动驾驶至关重要,为了设计适应多种环境更高效的自动驾驶控制算法,提出了基于强化学习的联合训练方法。自动驾驶控制算法应该在各种道路环境、各种天气环境、各种场景下都可以稳定的运行,因此通过人工智能的方法设计自动驾驶算法时必须要考虑到各种场景。基于TORCS仿真软件设计了基于强化学习的联合训练方法,包括使用神经网络拟合动作、状态空间,设置训练策略、奖励函数等机制。同时通过在不同环境下设置多个智能体进行强化学习训练,并设计了联合训练的算法,实现多个智能体在不同环境下进行联合训练,不同智能体共享相互学习到的经验,提高了模型的泛化性。所设计的联合训练方法实现了多个强化学习智能体的联合训练,并通过了实验验证,达到了高效、稳定的训练策略。

Automatic driving control algorithm is very important for automatic driving.In order to design more efficient automatic driving control algorithm adapted to various environments,this paper proposes a joint training method based on reinforcement learning.The automatic driving control algorithm should be able to run stably in various road environments,weather environments and scenarios.Therefore,to design an automatic driving algorithm through artificial intelligence,various scenarios must be taken into account.Based on TORCS simulation software,this paper designs a joint training method based on reinforcement learning,including using neural network to fit action and state space,setting training strategy,reward function and other mechanisms.At the same time,multiple agents are set up in different environments for reinforcement learning training,and the algorithm of joint training is designed to realize the joint training of multiple agents in different environments,and different agents share the experience learned from each other,which improves the generalization of the model.The joint training method designed in this paper realizes the joint training of multiple reinforcement learning agents,and is verified by experiments,and achieves an efficient and stable training strategy.