CollisionPro

KO | EN

gitlite — search

#javascript #ai #llm #go #machine-learning #rust #deep-learning #docker #java #windows #kubernetes #vue #awesome-list #ios

(★ 14)

Towards explainable value functions in reinforcement learning. A framework for collision probability distribution estimation via deep temporal difference learning.

CollisionPro 최신버젼 다운로드

최종 버전 다운로드 (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

<h1 align="center">
  <a href="https://ieeexplore.ieee.org/document/10920103">CollisionPro</a>
</h1>

`Towards explainable value functions in reinforcement learning——A framework for collision probability distribution estimation via temporal difference learning`

</div>

![Overview CollisionPro](https://raw.githubusercontent.com/UniBwTAS/CollisionPro/master/assets/overview_collisionpro.png "Overview CollisionPro")

---

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 read a comprehensive article about this work [here](https://medium.com/@steineckertommy/finite-horizon-value-function-43354b9405d9).  
The related research paper is available [here](https://arxiv.org/abs/2407.20000).

If you find our work helpful in your research, please consider citing our paper:

```
@INPROCEEDINGS{steinecker2024collision,
  author={Steinecker, Thomas and Luettel, Thorsten and Maehlisch, Mirko},
  booktitle={2024 IEEE 27th International Conference on Intelligent Transportation Systems (ITSC)}, 
  title={Collision Probability Distribution Estimation via Temporal Difference Learning}, 
  year={2024},
  pages={2645-2650},
  doi={10.1109/ITSC58415.2024.10920103}}
```

---

## Quick Links

+ [Installation](#installation)
+ [Run Examples](#run-examples)
+ [HowTo Guide](#howto-guide)

## Installation

To effortlessly install CollisionPro, utilize the following pip command:

```bash
pip3 install git+https://github.com/UniBwTAS/CollisionPro.git#egg=CollisionPro
```

or

```bash
python3 -m pip install git+https://github.com/UniBwTAS/CollisionPro.git#egg=CollisionPro
```

After a successful installation, ensure the correctness of the setup by following these steps:

1. Open a new terminal window
2. Launch the Python 3 shell:

```bash
python3
```

3. Verify the installation by importing the CollisionPro module:

```python
import collisionpro
```

## Run Examples

We 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](./docs/HowTo.md).

Once you have installed CollisionPro (see [Installation](#installation)), you can easily run our examples:

**Random Walk (Easy Example)**

![Random Walk environment with 7 states](https://raw.githubusercontent.com/UniBwTAS/CollisionPro/master/assets/docs/random_walk.png "Random Walk 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

```python
from collisionpro.examples.random_walk.main import run
run()
```

**Moving Circles (Advanced Example)**

![Moving Circles](https://raw.githubusercontent.com/UniBwTAS/CollisionPro/master/assets/moving_circles.png "Moving Circle Scheme")

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.

```python
from collisionpro.examples.moving_circles.main import run
run()
```

## HowTo Guide

Find a comprehensive Guide for CollisionPro functionalities and examples [here](./docs/HowTo.md).

CollisionPro

CollisionPro 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

CollisionPro

CollisionPro 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

Download README