data/results_plot/README.md

Final Results Plot

We use saved results by .npz type (Our algorithm, and CRM) and .json type (QRM). Some experiments include saved files as RLAlgorithm.zip, and saved results (success rate, mean reward, episode length, partial success) as .npz files.

• normal: This folder includes experimental results for a deterministic environment without noise, and infeasible conditions.

• noise01: This folder contains experimental results for a noisy environment without infeasible conditions.

• missing: This folder contains experimental results for an infeasible environment without noise.

• normal_csv: Contains .csv files to plot results in environments without noise and infeasible conditions.

• noise_csv: Contains .csv files to plot results in noisy environments without infeasible conditions.

• missing_csv: Contains .csv files to plot results in infeasible environments without noise.

Plotting Results:

1. Convert .npz and .json file results into .csv file results for each environment (deterministic, noisy, missing).

2. Use draw_with_csv.py to plot results for each environment.

3. Optional: To plot results on the HalfCheetah environment for the A2C and PPO algorithm, use draw_cheetah.py.

compute_mean_std.py

This file aims to compute the mean reward and standard deviation saved by npz files for the logic-based reward shaping algorithm and CRM, and the json file for QRM. Computed mean rewards and standard deviations will be saved as a csv file.

Arguments:

• First argument relates to the missing condition, which is boolean.
• Second argument relates to the noise condition and is a float. We've tested with a 0.1 (10%) chance of noisy control.

Example:

python compute_mean_std.py False 0

The example above converts saved results (either .npz or .json type) in a deterministic environment with a 0% chance of noisy control for various algorithms across different environments.

python compute_mean_std.py False 0

This example converts saved results of the A2C algorithm on the HalfCheetah domain.

draw_with_csv.py

This will read the converted csv file by compute_mean_std.py, then plot and save the results.

Arguments:

• First argument relates to type of plots (normal, noise, missing)

To see the result on deterministic environments without noise and infeaisble condition, Example:

python draw_with_csv.py normal

To see the result on environments with noise and without infeaisble condition, Example:

python draw_with_csv.py noise

To see the result on environments without noise and with infeaisble condition, Example:

python draw_with_csv.py missing

All plots will be saved ./saved_plots folder.

draw_cheetah.py

This is exclusively for the DDPG, A2C, PPO algorithm's results on the HalfCheetah Environment.

Once you have converted all saved results (.npz files) into CSV format with compute_mea_std.py, use

python draw_cheetah.py

All plots will be saved ./saved_plots folder.

plotter.py

This includes functions to aggregate all different types of results (QRM result is saved with .json, and the others are saved with .npz). These functions will not be used directly. Instead, these functions will be used when we convert .npz and .json format to CSV format with compute_mea_std.py.