This repository contains the code to replicate the experiment in the eponymous paper.
The data for the battery storage experiment (contained in battery_storage/Data/battery_storage) was copied from the respository associated with Donti et al. (2017). The original time series used to be published by PJM. We also borrow some of the data loading code in data_utils.py. The data for the portfolio optimisation experiment (contained in portfolio_optimisation/Data/portfolio_optimisation/CNNpred) was copied from repository.
 |
|---|
| Figure 1: Example model predictions on the battery storage task. |
The models are built with PyTorch and cvxpylayers, implementation is in models.py. We provide a requirements.py file to clarify other dependencies. To try the experiment with diffferent hyperparameters the hyperparameters.py file contains a function in which settings may be changed. To run the experiment first cd into the desired experiment and call run.py in which one can define how many replications to run. The file runs the experiment function defined in experiment_whole_policy_evaluation.py which contains the whole pipeline as assembled from other files.
- Results/policy/ contains the results of the experiments that we ran (not including saved models due to size in the case of portfolito optimisation).
- data_utils.py contains loading/processing functions for data and dataset classes for training and evaluation.
- evaluation_utils.py contains evaluation functions.
- optimisation_utils.py contains functions which parametrically build the cvxpy problem and optimisation layer, and the torch loss function for the DFF optimisation.