Multi-agent Lenient Reinforcement Learning Based Algorithm for Balanced Train Operation of Single-Track Railway Considering DoS Attacks

Abstract

In the actual two-way single-track rail transit system, the occupancy conflicts of the same station or the same track section will inevitably occur between the planned trains. Considering the influence of train velocity, the balanced train operation problem aims to find the optimal velocity of each train on the railway to minimize the delay ratio (that is, the ratio of the total delay time to the total free running time). However, with the expansion of the transportation network, there are hidden dangers in the security of centralized algorithms. In order to improve computing efficiency and security, distributed optimization methods are considered. In addition, due to the uncertainty of the train environment, there are usually uncertainties in the relationship between its velocity and delay time. Thus, a multiagent balanced train operation method is proposed, in which all agents make joint decisions by coordinating with the environment and local neighbors. Due to the introduction of distributed algorithms, network security also needs to be considered. The algorithm has designed a defense mechanism against Dos attacks. Due to the uncertainty of the cost functions, lenient method is adopted to estimate the rewards with neural networks. Through the case studies, the superiority of the algorithm is verified.