Learning to Drive in the NGSIM Simulator Using Proximal Policy Optimization

As a popular research field, autonomous driving may offer great benefits for human society. To achieve that, current studies often applied machine learning methods like reinforcement to enable an agent interact and learn in stimulating environment. However, most simulators lack realistic traffic which cause deficiency interaction. The present study adopted the SMARTS platform create simulator trajectories of vehicles NGSIM I-80 dataset were extracted as background traffic. built was used train model using proximal policy optimization method. actor-critic neural network applied, takes inputs including 38 features that encode information host vehicle nearest surrounding lane adjacent lane. A2C selected comparative results revealed PPO outperformed task by collecting more rewards, traveling longer distances, encountering less dangerous events during training testing. achieved 84% success rate test is comparable related studies. proved public can provide useful tool driving.