MetroGNN: Metro Network Expansion with Deep Reinforcement Learning
Primary Area: applications to robotics, autonomy, planning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: metro network expansion, reinforcement learning, graph neural networks
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
TL;DR: We propose a graph-based reinforcement learning framework to solve complex node selection MDP on the graph, which achieves superior performance in the metro network expansion task.
Abstract: Selecting urban regions for metro network expansion that serve maximal transportation demands is critical to urban development, while computationally challenging to solve. First, metro network expansion is dependent on multiple complicated features, such as urban demographics, origin-destination (OD) flow, and relationships with existing metro lines, requiring a unified model to incorporate these correlated features for region selection. Second, it is a complex decision-making task with an enormous solution space and various constraints, due to the large number of candidate regions and restrictions on urban geography. In this paper, we present a reinforcement learning framework to solve a Markov decision process on an urban heterogeneous multi-graph, achieving metro network expansion by intelligently selecting a set of nodes on the graph. A novel graph neural network is proposed, which unifies the complicated features and learns effective representations for urban regions. In addition, we design an attentive reinforcement learning agent with action masks to efficiently search the large solution space and avoid infeasible solutions indicated by the various constraints. Experiments on real-world urban data of Beijing and Changsha show that our proposed approach can improve the satisfied transportation demands substantially by over 30\% compared with state-of-the-art reinforcement learning methods. Further in-depth analysis demonstrates that MetroGNN can provide explainable results in scenarios with much more complicated initial conditions and expansion requirements, indicating its applicability in real-world metro network design tasks. Codes are released at https://anonymous.4open.science/r/MetroGNN-31DD.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 7259
Loading