LPAC Neural Network

Table of Contents

• Preliminaries
• Dataset Generation
• Training
• Evaluation

Preliminaries

We will organize files in a workspace directory: ${CoverageControl_ws} (e.g., ~/CoverageControl_ws).

Download and extract the file lpac_CoverageControl.tar.gz to the workspace directory. The file can be downloaded from the repository releases.

tar -xvzf lpac_CoverageControl.tar.gz -C ${CoverageControl_ws}

This will create a directory lpac in the workspace directory. The directory structure is as follows:

${CoverageControl_ws}/
└── lpac/
    ├── data/   # To store datasets
    ├── envs/   # Environment files
    ├── eval/   # Results of evaluation
    ├── models/ # Trained models
    └── params/ # Parameters for training and evaluation

The models folder already contains a trained LPAC model for a 1024x1024 environment with 32 robots, 32 features, and 128 communication radius.

Dataset Generation

There are two ways to classes for dataset generation located in python/data_generation/

1. simple_data_generation.py
2. data_generation.py

They are similar, except that data_generation.py splits the dataset into training, validation, and test sets.

To generate a dataset, run the following command:

python python/data_generation/data_generation.py \
       ${CoverageControl_ws}/lpac/params/data_params.toml

A sample data_params.toml file is also provided in the params directory of the repository. See the file for details on the parameters. The class will use a coverage_control_params.toml configuration file to generate environments and then use the ClairvoyantCVT algorithm to generate the dataset.

The simple_data_generation.py is useful for generating a large dataset in parts and then combining them into a single dataset. See python/utils/process_data.sh and python/utils/dataset_utils.py for tools to process and combine datasets.

Training

To train the LPAC model, run the following command:

python python/training/train_lpac.py \
       ${CoverageControl_ws}/lpac/params/learning_params.toml 1024

The second argument is the environment size, used to normalize the input features. A sample learning_params.toml file is also provided in the params directory of the repository. See the file for details on the parameters.

Evaluation

There are two scripts for evaluation located in python/evaluators/

1. eval_single_env.py
2. eval.py

eval_single_env.py evaluates a single environment and eval.py evaluates multiple environments.

To evaluate a trained model, run the following command:

python python/evaluators/eval.py \
       ${CoverageControl_ws}/lpac/params/eval.toml

or

python python/evaluators/eval_single_env.py \
       ${CoverageControl_ws}/lpac/params/eval_single.toml

The eval.toml and eval_single.toml files are also provided in the params directory of the repository.