Implementing Deep Reinforcement Learning (DRL)-based Driving Styles for Non-Player Vehicles

A realistic representation of the traffic participants is a key feature of serious games for driving. We propose a novel method for improving the behavioral planning of non-player vehicles (NPVs) by supporting implementation of (i) “human-like”, high-level decision making and (ii) different driving styles. The method relies on the deep reinforcement learning (DRL) technology, which is gaining ever more interest in the real-world automated driving research. We designed a system architecture including advanced driving assistance systems (ADAS) and trained the agent in the highway-env DRL environment. Compared to a low-level decision making system, our system performs better both in terms of safety and speed. Moreover, the proposed approach allows reducing the number of training steps by more than one order of magnitude. This makes the development of new models much more efficient. As a second main contribution, we demonstrate that it is possible to train agents characterized by different driving behaviors, by tweaking the weights of the factors of a general DRL reward function. This approach avoids heuristic coding of the driving styles, saving development and maintenance time. The developed agent models can then be seamlessly deployed as NPVs in any target SG in the same development environment (i.e., highway-env). Furthermore, the information and lessons learned presented in this article can be useful to effectively and efficiently train similar agents in their target SG environment.