LaGCO-RL: Reinforcement Learning in Latent Action Spaces for Graph Combinatorial Optimization
Year: 2026
Type: software
Abstract: LaGCO-RL is a modular library that automates the construction of latent action spaces and semantically aligned latent observation spaces for reinforcement learning (RL) in graph-based combinatorial optimization (GCO) through graph neural network (GNN) embeddings.The framework supports a diverse set of GCO benchmark problems, ranging from tasks with linear node decision variables to more realistic scenarios involving structured and super-linear decision variables. This broad applicability addresses the action-space complexity gap present in existing GCO benchmarks. LaGCO-RL is also easily extensible, enabling the integration of new optimization tasks by defining only the environment dynamics and feature extraction pipeline.In addition, LaGCO-RL provides implementations of multiple RL-based approaches for GCO, including standard discrete-action methods, iterative Q-learning agents, and the projection-based RL agent proposed in this work.
Source: HAL
Institutions: Institut national de recherche en sciences et technologies du numérique, Laboratoire Lorrain de Recherche en Informatique et ses Applications, RESIST: Résilience et élasticité pour la mise à l'échelle et la sécurité des systèmes en réseaux dynamiques, Université de Lorraine, University of California, Santa Barbara, Universitat Politècnica de Catalunya
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7170236310