Optimization-Based Algebraic Multigrid Coarsening Using Reinforcement Learning

Code for reproducing the experimental results in our paper: https://arxiv.org/pdf/2106.01854.pdf

Requirements

• Python = 3.8.10
• PyTorch = 1.9.0
• Pytorch_Geometric = 1.7.2
• pyamg = 4.1.0
• networkx = 2.6.2
• pygmsh = 7.1.6
• gmsh = 4.8.0

Training

python main.py

Files

• Batch_Graph.py: The class for making a batch of grids to be the input to the agent's network.
• DuelingNet.py: The netwrok architecture that the agent utilizes (TAGCN network).
• Lloyd_Unstructured.py: The code for the generating Lloyd agregations on grids.
• MG_Agent.py: The class for Dueling Double DQN agent.
• Optim.py: The code for running the grid coarsinging using Lloyed aggregation.
• Solve_MG.py: The code for the two grid cycle AMG algorithm.
• Unstructured.py: The class for defining grids and making unstructured meshes.
• fem.py: Constructs a finite element discretization of a 2D Poisson problem.
• main.py: The main driver for training the RL agent and test functions.