abstract
Through this project, we aim to test the applicability of Multi-Agent methods to the efficient automated scheduling of trains. To this end, we chose the Flatland 3 challenge, a multi-agent reinforcement learning environment that simulates a railway network, in order to have a setting realistically similar to the real-world problem. We implemented a Multi-Agent Proximal Policy Optimization (MAPPO) algorithm, together with two observations tailored
to the problem and two network architectures. We evaluated our trained models and found that our best model would have reached the 1st place on the RL track of the Flatland 3 challenge.
outcomes
