Towards XAI
A framework for collision probability distribution estimation via temporal difference learning
In today's deep learning era, understanding the inner workings of complex models is of major interest. Enter CollisionPro, a noval framework designed to estimate cumulative collision probability distributions through temporal difference learning.
But what sets CollisionPro apart? It's all about transparency and explainability. In an age where black-box algorithms reign supreme, CollisionPro takes a step towards explainable AI. By generating a cumulative probability distribution, it offers not just a single risk assessment value, but a spectrum of interpretable values. Moreover, each value could be interpreted in the context of reinforcement learning.
You can find the paper of this work here.
If you find our work useful in your research, please consider citing our paper:
@article{collisionpro,
title={Collision Probability Distribution Estimation via Temporal Difference Learning},
author={Thomas Steinecker and Thorsten Luettel and Mirko Maehlisch},
journal={arXiv preprint arXiv:2407.20000},
year={2024},
}
To effortlessly install CollisionPro, utilize the following pip command:
pip3 install git+https://github.com/UniBwTAS/CollisionPro.git#egg=CollisionProor
python3 -m pip install git+https://github.com/UniBwTAS/CollisionPro.git#egg=CollisionProAfter a successful installation, ensure the correctness of the setup by following these steps:
- Open a new terminal window
- Launch the Python 3 shell:
python3- Verify the installation by importing the CollisionPro module:
import collisionproWe provide some examples for an easier/faster understanding of CollisionPro. Each example consists of an environment (env.py), an approximator (approximator.py), that learns the probability distribution, a controller (controller.py) if required, that provides an action given the current state, and a main script (main.py) that is setting up the training pipeline. For more information see HowTo Guide.
Once you have installed CollisionPro (see Installation), you can easily run our examples:
Random Walk (Easy Example)
The Random Walk environment consists of sequential states, where the first and last states indicate terminal states. The agent transitions with a 0.5 probability to the right and a 0.5 probability to the left. For each non-terminal state, the agent receives a reward of r=0, while for terminal states, the agent receives a reward of r=-1, which corresponds to a collision.
- No action space → no controller
- No value function approximation
from collisionpro.examples.random_walk.main import run
run()Moving Circles (Advanced Example)
The Moving Circles environment is centered around the ego circle (blue), which moves along the x-axis with a constant linear velocity. Obstacles (red circles) oscillate along the y-axis and can be modeled as mass-spring systems (no energy dissipation). When the outer circles of the ego and any obstacle overlap, a collision event is triggered, terminating the episode and resulting in a reward of -1.
Both the ego and obstacles predict each other's movements using constant kinematics into the future and adjust their positions accordingly. The ego ball can move along the y-axis and is modeled in y-direction as a mass-spring-damper system, while the obstacles can move along the x-axis.
- Action space available → controller provided
- Value function approximation via TensorFlow
- Multi-agent, partially observable, continuous environment
Execute the following, which will iteratively generate samples, learn the collision probability distribution and finally generates collision characteristics that are plotted.
from collisionpro.examples.moving_circles.main import run
run()Find a comprehensive Guide for CollisionPro functionalities and examples here.