Research on intelligent algorithm for alerting vehicle impact based on multi-agent deep reinforcement learning

Abstract

In previous work, we designed a smart wearable device and algorithm for early warning of dangerous vehicles for pedestrians. The algorithm uses the fuzzy comprehensive evaluation as the framework, and calculates the risk degree using the vehicle, environment, and pedestrian status as evaluation indicators. Then according to the different environment, the back-propagation neural network (BP neural network) is used to assign dynamic weights to the indicators. Finally, the overall risk is obtained by calculating the product of the risk degree vector and the weight vector. However, due to the narrowness of human experience and the insufficient learning samples, the neural network overfits on the training set and performs poorly on the test set. In the current work, we try to apply reinforcement learning (RL) and multi-agent reinforcement learning (MARL) to drive agents to generate samples, and use this to replace learning samples generated by expert experience. In addition, the deep Q-learning network (DQN) mechanism is applied in the agent training method to avoid the curse of dimensionality, and finally multi-agent deep reinforcement learning (MDRL) is applied. Different from the previous algorithm, this algorithm integrates the fuzzy calculation process that must be completed on MATLAB into the neural network during training, and thus can greatly improve the calculation speed. Through the simulation experiments using python, it proves that this method can improve the accuracy of the algorithm up to about 96% and perform the best delay level.