The value of AI signal control is that the green timing is learned by computer itself and without any pre-given expert system rule. The advantage of learning by itself is that control strategy is no longer dependent on specific traffic model. This feature is called "Model-Free".
The characteristics of Model-Free can be explained as the compatibility of different traffic model. For example, the same set of AI learning methods can be applied from macroscopic traffic flow models to microscopic traffic flow models, or even actual traffic environment data. That is to say, AI learning method can optimize the control strategy no matter how special traffic behavior the field is.
Traditional adaptive control usually link traffic model with it's decision logic. The adaptive control must cooperate with corrections or even unusable if the traffic flow model cannot accurately reflect the actual traffic state, or the decision logic cannot apply to the on-site traffic state. However, thanks to the characteristics of Model Free, the flexibility and scalability of AI method have a breakthrough improvement compared with traditional methods.
For the AI signal control system, there is no need to manually segment different operating periods or set different parameters in advance. The trained AI control logic can instantly calculate the best signal timing (each green time) in different traffic conditions to achieve the overall system performance optimization.
The AI signal control system is based on a machine learning method. It trains the AI signal control logic with the goal of optimizing the overall road network system. Through dynamic coordination between intersections, it achieves continuation of arterial roads and optimization of the road network.
The training method of AI signal control technology is universal and does not need to be redesigned for different fields. It is quite easy to expand the scope of control or transplant to other fields in the future.
AI signal control technology can deal with the problem of mixed traffic flow, e.g. scooter and passenger car in the same lane. Further, special traffic operating rules, like two-stage left turn for scooter, can be Included in model among the considerations.
Compared to the traditional method that converts scooter to PCU, AI traffic control does not need to do such man-made definition, and can compute green time by consider the behavior of scooter. Therefore, AI traffic control is more suitable to deal mixed traffic flow then traditional way.
Under the machine learning architecture, when the traffic demand shifts after a long time from now, the AI signal control system can still gain feedback form the experience, which is interacting with the actual environment, and keep learning and update the signal control strategy to catch up the changing of traffic patterns.
During training process, we will set many different simulation scenario to test the AI's reaction. Through various scenario, we can find the weakness and then strengthen before practical.
In addition, system will have degraded operation function (fallback routine), when accident happened, e.g. failure at detection equipment, communication equipment, control equipment.