Projection Agents: A New RL-GCO Approach for Graph Optimization
A new arXiv preprint (2605.19721) introduces projection agents for graph combinatorial optimization (GCO). The method operates in a continuous GNN-based action embedding space, predicting latent actions in a single forward pass and decoding them into valid discrete actions. It addresses generalization and scalability issues in existing RL-GCO solvers by enabling fair comparison across methods via a shared embedding space for observations and actions. The approach is evaluated on diverse benchmarks.
Key facts
- arXiv:2605.19721
- Projection agents operate in continuous GNN-based action embedding space
- Predicts latent action in single forward pass
- Decodes latent action into valid discrete action
- Enables fair comparison across RL methods via shared embedding space
- Addresses generalization and scalability challenges in GCO
- Evaluated on diverse benchmarks
- Combines Reinforcement Learning with Graph Neural Networks
Entities
Institutions
- arXiv