我们研究通过强化学习来使用深度Q学习控制模拟自主车辆。我们首先实现了文献[5]提出的方法，然后尝试各种可能的更改，以提高所选任务的性能。特别是，我们实验了各种奖励函数来诱导特定的驾驶行为、双Q学习、梯度更新规则和其他超参数。我们发现在某些方面能够成功地训练一个代理来控制Javascript Racer [3]中的模拟车辆。我们的智能体成功地学会了转弯操作，逐步获得了在模拟赛道上长时间行驶而不撞车的能力。然而，在障碍规避方面，仍然面临着挑战，我们怀疑这是由于训练时间不足的原因。

We investigate the use of Deep Q-Learningto control a simulated car via reinforcement learning. We start by implementingthe approach of [5] ourselves, and then experimenting with various possiblealterations to improve performance on our selected task. In particular, weexperiment with various reward functions to induce specific driving behavior,double Q-learning, gradient update rules, and other hyperparameters. We find weare successfully able to train an agent to control the simulated car inJavaScript Racer [3] in some respects. Our agent successfully learned theturning operation, progressively gaining the ability to navigate largersections of the simulated raceway without crashing. In obstacle avoidance,however, our agent faced challenges which we suspect are due to insufficienttraining time.

更多精彩文章请关注公众号：