Rothe_L2O

In this project, we implement Learning to Optimize (L2O) to learn a LSTMs and n-SSMs, that aim to optimize

1. Quadratic functions (as a test). This is done using LSTMS ans n-SSMs. See the directory L2O_quadratic.
2. The Rothe error for the Hydrogen atom at different field strengths, time steps, and wave function qualities. This was done using LSTMs. See the directory L2O_Hydrogen.

Furthermore, we have implemented a T-GCN to predict the best starting guess for the nonlinear coefficients at the next time step for the Henon-Heiles model in 3D. See the directory tGCN-Henon_Heiles.

Usage

1. For the test problems, run the models (L2O_quadratic_singleVariable.py, ssm_l2o.py and s4_ssm.py), then the respective run_L2O_experiment_"model".py, before finally running plot_L2O_experiment_"model".py. Each "model".py takes command line arguments T (unrolling depth) and k (weight scaling). The experiement codes take three command lines T, k and type ('training' or 'testing'). The plotting codes, similar to the model codes, take T and k as command line arguments.

   ssm_l2o.py 20 1
   run_L2O_experiment_SSM.py 20 1 'training'
   run_L2O_experiment_SSM.py 20 1 'testing'
   plot_L2O_experiment_SSM.py 20 1
   

2. For L2O_Hydrogen, run L2O_hydrogen.py narrow and L2O_hydrogen.py wide for the sweep over all metaparameters. To find the configuration for the best metaprameters, run python3 analyze_runs.py narrow and python3 analyze_runs.py wide. Run train_and_evaluate_best_model.py to re-train and re-evaluate the best models. Finally, run make_plots.py narrow and make_plots.py wide to reproduce the plots.

3. For the t-GCN, run train_models.py to do a sweep over the metaparameters. Run find_best_values.py 5 True, find_best_values.py 5 False, find_best_values.py 10 True, find_best_values.py 10 False to find the best metarparameters & epoch. This also re-trains the best model and produces an output. Run create_results.py 5 True, create_results.py 10 True, create_results.py 5 False, create_results.py 10 False, to create the MSE data that is visualized in the paper. Finally, run plot_results True and plot_results False to reproduce the plots.