Part III Project: "Generalisable data-driven routing using RL with GNNs"

This is the code accompanying my dissertation submitted for my masters degree (which can be found at https://github.com/odnh/gnn-routing-diss).

The aim is to extend work done in "Learning to route with deep RL" to use GNNs, hopefully allowing generalisation to unseen graphs (read the dissertation for further details and a more in depth explanation).

Layout:

• gym-ddr: Implementation of openai gym env for RL routing
• stable-baselines-extensions: modifications to stable-baselines for custom learning policies using GNNs.
• dd-learning-helpers: library of helper functions that don't rely on gym or stable-baselines
• experiments: scripts to learn and test models
• data: contains data to use in experiments (e.g. graphs, demands)
• evaluation: contains scripts that combined with a configuration will run the experiments for project evaluation and plot the results.
• raeke: OCaml code for generating an oblivious routing. Has issues so was not used.

Requirements:

Based on stable-baselines and using Graph Nets for GNN bits and pieces. To set up on a Linux (Ubuntu) machine from scratch, run the bootstrap.sh file (Of course, read what the script does before running it).

To setup python env, env.yml can be used with conda env -f env.yml