yufeiwang63 / L2D-learning-to-discretize

source code of learning to discretize solving 1d scalar conservation laws via deep reinforcement learning

Learning to Discretize: Solving 1D Scalar Conservation Laws via Deep Reinforcement Learning

This is the code for the paper "Learning to Discretize: Solving 1D Scalar Conservation Laws via Deep Reinforcement Learning", published at Communication in Computational Physics (CiCP) 2020. [Arxiv]. Authors: Yufei Wang*, Ziju Shen*, Zichao Long, Bin Dong (* indicates equal contribution)

Instructions

Preparation

1. Create conda environment

conda env create -f environment.yml

2. Activate the conda environment

source prepare.sh

Training

1. Generate training datasets

python scripts/generate_solutions.py --prefix [PREFIX] --eta [ETA] --num [NUM]

to generate and store solutions to some randomly generated initial conditions. The solution files will be stored at data/local/solutions/prefix

2. Perform training

python launchers/launch_td3.py

This will train a TD3 agent on the pre-generated solutions of different initial conditions in step 1.
Remember to change the solution_data_path variable to point to the correct path of the generated solutions in step 1.
You can tune the parameters in launch_td3.py.

The training logs will be stored at data/local/exp_prefix. exp_prefix is specified in launch_td3.py.
To view the training logs, you can use python viskit/frontend.py data/local/exp_prefix.
To make videos of the trained models, use scripts/make_videos.py
To make error tables of the trained models, use scripts/test_trained_model_mesh.py or scripts/test_trained_model_forcing.py or scripts/test_trained_model_viscous.py