Skip to content

Latest commit

 

History

History
96 lines (87 loc) · 5.84 KB

README.md

File metadata and controls

96 lines (87 loc) · 5.84 KB
Raw

kf_bayesopt

Kalman Filter tuning example using Bayesian Optimization. If you use the relates code, please consider citing

@inproceedings{chen2018weak,
  title={Weak in the NEES?: Auto-tuning Kalman filters with Bayesian optimization},
  author={Chen, Zhaozhong and Heckman, Christoffer and Julier, Simon and Ahmed, Nisar},
  booktitle={2018 21st International Conference on Information Fusion (FUSION)},
  pages={1072--1079},
  year={2018},
  organization={IEEE}
}

@article{chen2021time,
  title={Time Dependence in Kalman Filter Tuning},
  author={Chen, Zhaozhong and Heckman, Christoffer and Julier, Simon and Ahmed, Nisar},
  journal={arXiv preprint arXiv:2108.10712},
  year={2021}
}

Fusion 2021 has accepted the second reference.

Introduction

This package simulates 1D robot with linear kinematics model as decribed in our paper <Weak in the NEES?: Auto-tuning Kalman Filters with Bayesian Optimization>

This system is also build on ROS so you also need install ROS. With the help of ROS, bayesian optimization lib and robot1d lib can communicate with each other more easily. At the same time, all the parameters of bayesian optimization and robot1d can be modified in the yaml file. The main structure can be seen from the following figure image

robot1d_KF node will run simulator and estimator(KF) accodring to different noise setting. It will output J_NEES or J_NIS(topic "cost") and publish to the robot1d_bayesopt node. Acoording to the cost, robot1d_bayesopt will run the bayesian optimization to optimize the cost, then it will generate the noise(topic "noise") to publish to the robot1d_KF and set the nosie. Then robot1d_KF node can run the simulator and estimator agian and generate new cost. To modify the parameters, please go to the config. There are two yaml files bayesParams.yaml to modify the parameters of bayesopt and vehicleParams to modify the parameters of robot1d.

Dependencies

Install bayesian optimization

bayesian optimization is from my github https://github.com/zhaozhongch/bayesopt. (The official bayesopt will also work but no API for python and matlab plot)

sudo apt-get install libboost-dev python-dev python-numpy cmake cmake-curses-gui g++ cython octave-headers freeglut3-dev
git clone https://github.com/zhaozhongch/bayesopt.git
cd bayesopt
mkdir build
cd build
cmake ..
make
sudo make install

Deatails can be seen at https://rmcantin.bitbucket.io/html/install.html

ROS indigo or kinect or melodic

Eigen3

Install the bayesopt_ros package

mkdir -p catkin_ws/src
cd catkin_ws
catkin_make
cd src
git clone --single-branch -b master https://github.com/arpg/kf_bayesopt.git
cd ..
catkin_make

Run the code

Run bayesian optimization for robot1d problem
open one terminal

source ~/catkin_ws/devel/setup.bash
roslaunch robot_kf_bayesopt robot1d_kf.launch

open another terminal

source ~/catkin_ws/devel/setup.bash
roslaunch robot_kf_bayesopt robot1d_bayesopt.launch

In 1D or 2D case, if you download the bayesopt from my github, you can visualize the process in real time, open another terminal

source ~/catkin_ws/devel/setup.bash
rosrun robot_kf_bayesopt plot_surrogate_acquisition.py

For 1D optimization, you'll see something like this image In the example image, the upper one is surrogate model with mean, 95% upper bound, 95% lower bound and sample points. The lower one is acquisition function. you can see the legend to know the detail.
For 2D optimization, you'll see something like this image In this example image, you'll see 2d surrogate model and 2d acquisition function. No upper and lower bound is shown because the bound will cover the mean so I decide not to add them.
What's more, we also provide matlab script to plot the bayesopt for 1D and 2D problem. For 1D and 2D optimization, every time the optimization is finished, you should be able to see some csv files in the bayesopt/bayesplot folder. With the matlab script, specify the location of the csv folder, output image folder and initial sample number, for 1D/2D optimization, you can get a plot like the following image image

Parameters

The parameter of bayesopt should be considered from the bayesopt official website https://github.com/rmcantin/bayesopt. You can specify those parameters in bayesParams.yaml. The parameters of the simulation robot are pretty strightforward if you read the paper and the parameters are stored in vehicleParams.yaml
One thing you need to notice is that the upper bound and lower bound's dimension in bayesParams.yaml must agree with the optimizationChoice parameter in vehicleParams.yaml.
For example, the processnoise is 1d, if you choose optimizationChoice to be processNoise then the lower and upper bound should be just 1 dimension; if you choose optimizationChoice to be all then the lower and upper bound should be just 2 dimension.
CPUcoreNumber decides how many threads you wan to run parallely, which you can choose according to the core number of your computer.

Two sample time approaches

We found the cost function is ambiguous and has observability problems. The sample time is a key component in our solution to relieve the problem. For each set of noise parameters, we run the Kalman filter twice with two different sample time (no strict requirement as long as the two sample time are different enough. For example 0.1 and 0.5). We pick the larger cost and feed it into the optimization. The reason of this solution can be seen in our second reference chen2021time.