T2.3_explainability_dashboard

This repository provides an explainability panel for RL agents developed for Power Grid use case. It is related to Task 2.3 of AI4REALNET project.

Installation guide

To be able to run the experiments in this repository, the following steps show how to install this package and its dependencies from source.

Setup a Virtualenv (optional)

Create an environment using Conda (recommended)

conda create -n env_explainability python=3.10
conda activate env_explainability

Create a virtual environment using virtualenv

cd my-project-folder
pip3 install -U virtualenv
python3 -m virtualenv my_env
source my_env/bin/activate

Install the prerequisites

Important

These steps are mandatory to be able to use the package and its different functionalities

ExpertOp4Grid package

git clone git@github.com:Mleyliabadi/ExpertOp4Grid.git
cd ExpertOp4Grid
pip install -U .

LJN Agent package

git clone git@github.com:Mleyliabadi/l2rpn-2023-ljn-agent.git
cd l2rpn-2023-ljn-agent
pip install -U .

ExpertAgent package

git clone git@github.com:AI4REALNET/T2.1_deep_expert.git
cd T2.1_deep_expert
pip install -U .

Install the requirements for the current package

pip install -r requiremnets.txt

Requirements

Two major requirements for this explainability panel are:

1. Agent: An already trained agent on one of the power grid environments

Warning

This package works only with ExpertAgent Heuristic developed in the context of AI4REALNET project. It could be improved in future to be adapted to other agents.

2. Environment: One of the power grid environments designed for the projet. It could also be used for other environments.

Warning

This package works only with ai4realnet_small environment designed for AI4REALNET project and that corresponds to the first usecase of Power Grids. It could be improved in future to include other environments.

3. ExpertOp4Grid: You should configure the ExpertOp4Grid configuration file, which is available for an example in explainability_panel/config_expert.ini

Dashboard

A screenshot of the Explainability dashboard is provided below. It composed of two main parts:

1. Statistics Panel: some statistics allowing to analyze in more detail the agent's action and its impacts. These statistics are:

   • The number of hops that the agent's action is distant from the expert interest zone;
   • How much the agent's action allows to reduce the maximum overload on the grid (which is expressed by $\rho$);
   • The difference between the $\rho$ values affected by the action of the actual agent and the best action proposed by the expert;
   • The Wilcoxon test with $95%$ of confidence between the top-k agent's action distribution ranks and their corresponding ranks when using a greedy strategy. If $p-value$ is less that $0.05$, we reject the null hypothesis telling that there is no significant difference between the agent's action and best actions found using the greedy strategy. Hence, the agent's action may not be reliable.

2. Visualization panel: showing whether the grid or an overload graph (in the case of overload) obtained using the ExpertOp4Grid package and highlighting the highest impact zone near the overload as well as the actual agent's action. The agent recommendations will also be shown on this graph.

Usage

You can run the dashboard by simply executing the following command:

cd explainability_panel
python explainability.py

For the moment, an agent is hard coded inside the code. Future developments will allow the user to import a custom agent from the interface and analyze its recommendations.

There are three buttons:

1. Load Env: Load (reset) the environment and load the agent.
2. Next scenario: One step of the environment.
3. Recommend Action: Recommend an action from the considered agent.

Two addtional zones:

• Explanation: A text interpreting the statistics shown in the left side of the panel.
• Action description: prints the action remmended by the Agent.

License information

This Source Code is subject to the terms of the Mozilla Public License (MPL) v2 also available here