[Feat] Implementing GLOP

[Furffico]

Description

Following #182, this PR includes the implementation of Global and Local Optimization Policies (GLOP) (Ye et al., 2023), along with these new features for reproducing the results:

• New model: GLOP (based on the implementation in [Feat] Adding GLOP model #182)
• New environments:
  • Shortest Hamiltonian Path Problem (SHPP): Full credit goes to @cbhua [Feat] Adding GLOP model #182
  • CVRP with Maximum Vehicle Constraint
• New embeddings: Init/Edge embeddings based on polar coordinates.
• Two adapters for decomposing and composing TSP/CVRP solutions.

Types of changes

• New feature (non-breaking change which adds core functionality)
• Documentation (update in the documentation)

Checklist

• My change requires a change to the documentation.
• I have updated the tests accordingly (required for a bug fix or a new feature).
• I have updated the documentation accordingly.

CC: @henry-yeh @fedebotu

[Furffico]

Despite all these efforts, the current version of the GLOP model still refuses to learn 😞. However, I've obtained a working checkpoint in Oct. 2024. Since then, I’ve tried numerous times to replicate the success, but no luck so far. It seems like getting it to learn requires a bit of magic 🤔. At least this proves that the basic code logic is working as expected.

In theory, this checkpoint should work with this version of code. I’ll upload the checkpoint soon after some tests!

[fedebotu]

Despite all these efforts, the current version of the GLOP model still refuses to learn 😞. However, I've obtained a working checkpoint in Oct. 2024. Since then, I’ve tried numerous times to replicate the success, but no luck so far. It seems like getting it to learn requires a bit of magic 🤔. At least this proves that the basic code logic is working as expected.

In theory, this checkpoint should work with this version of code. I’ll upload the checkpoint soon after some tests!

Great job @Furffico !

I think it may also be a problem of the environment in reproducing the main paper results, as seen here. Perhaps downgrading to e.g. PyTorch 2.2 / changing to FP32 could do the trick?

This said, given the code is indeed correct per se if it's indeed a matter of environment as it seems to be I would be for merging!