Multi-agent reinforcement learning for shared resource scheduling conflict resolution

Transportation operations especially in railroad domain are time critical. Scheduling conflicts driven by disruptions and delays in any one zone significantly affect the overall network operations. In this work applicability of multi agent reinforcement learning approach to resolve scheduling conflicts and improve the railroad network operations was explored. Based on a custom 2D grid environment here we attempt to learn ideal coordinated agent actions based on simulated schedule conflict by introducing stochastic delays in train arrival. We were able to achieve converges for multi-agent simulation based setup with 30% malfunction rate. The focus of work is to presents the problem setup in mobility domain and simulation design for the multi-agent reinforcement learning. With respect to real world application this approach is promising as it reduces the requirement of a highly customized solution by experts and if a high-performance simulation-based reinforcement learning solution is reached this would provide an opportunity to build generalized interoperable control techniques for transit systems across the world.