WrapperRL: Reinforcement Learning Agent for Feature Selection in High-Dimensional Industrial Data

This code provides the implementation of the WrapperRL framework that emulates the wrapper-based method for forward feature selection using an actor-critic Reinforcement Learning (RL) algorithm. This method is applied to image-based high-dimensional time-frequency domain data collected in an industrial environment. The main goal of WrapperRL is to identify the set of features or image patches, representing a set of frequency bands, that are affected by ambient industrial noise. In addition to the implementation details of WrapperRL, the code includes the implementation of the SOTA approach PatchDrop and the ablation study.

The code includes the following folders:

• src: is the directory that includes the WrapperRL code.
• dataset: is the directory that includes the dataset used for building the models and obtaining the paper's results (img_dataset.csv).
• models: the directory that includes the classification model termed SA-CNN.pt.
• notebooks: the directory that includes the code for training the SA-CNN model.

To run the corresponding codes, the following steps need to be applied:

• The image dataset can be downloaded using the open-source link https://drive.google.com/file/d/1h4MgVdPRQxdc8pV51ZiANcnu_M3MFThl/view?usp=sharing. Please download the dataset and put it into the generated_images folder.
• Alter the IMGS_DIR variable in src/CNN_model/constants.py to the absolute path of the generated_images directory. An example would be r"D:\WrapperRL_AAAI\datasets\generated_images\log_all_features".

Code structure Explanation:

The codes in (1) src/main_grid.py, (2) src/main_patchDrop.py, and (3) src/main_randomPatch.py follow the same structure in terms of executing a forward feature selection process to our use case.

• The code in (1) runs the WrapperRL process. In this file, the trial_nb variable represents the trial number, w represents the lambda variable of the reward function, and total_episodes represents the training epochs. During the training process, an insights folder is created with the corresponding trial_nb and w to store the generated models, images, statistics, and training and testing datasets for a fair assessment.
• The code in (2) runs the PatchDrop process. In a similar manner, an insights folder is created to gather information about the process.
• The code in (3) runs the ablation study.

The codes in (1) src/analyze_results.py and (2) src/analyze_patchDrop.py analyzes the obtained results.

Methodology

Contact-Info

Please feel free to contact me for any questions or research opportunities.

• Email: [email protected]
• Gihub: https://github.com/ibrahimshaer and https://github.com/Western-OC2-Lab
• LinkedIn: Ibrahim Shaer
• Google Scholar: Ibrahim Shaer and OC2 Lab