This is the official code repository of the paper Lagrangian Reachtubes: The next Generation published at Proceedings of the 59th IEEE Conference on Decision and Control (CDC), 2020 (arXiv link). This paper received the IEEE award TC-DES and TC-HS Outstanding Student Paper Prize.
LRT-NG is a toolset that computes a reachtube (an over-approximation of the set of reachable states over a given time horizon) of continuous-depth neural networks as well as any other nonlinear dynamical systems.
This document describes the general usage of LRT-NG.
Click to expand and see the requirements and installation of LRT-NG
Download file from Ibex, create ibex directory, unpack downloaded file there, configure and install it:
curl http://www.ibex-lib.org/ibex-2.7.4.tgz --output ibex-2.7.4.tar.gz
tar xvfz ibex-2.7.4.tar.gz
cd ibex-2.7.4
sudo ./waf configure --interval-lib=filib
sudo ./waf install
Eigen is a C++ template library for linear algebra: matrices, vectors, numerical solvers, and related algorithms.
If you just want to use Eigen, you can use the header files right away. There is no binary library to link to, and no configured header file. Eigen is a pure template library defined in the headers.
git clone https://gitlab.com/libeigen/eigen.git /usr/local/include/eigen3
The program_options library allows program developers to obtain program options, that is (name, value) pairs from the user, via conventional methods such as command line and config file.
Download via package manager:
Linux
sudo apt-get install libboost-all-dev
Mac Os X using Homebrew
brew install boost
Download via package manager:
Linux
sudo apt-get install cmake
Mac Os X using Homebrew
brew install cmake
Clone this repository and use cmake to compile the code.
cd ~/
git clone [email protected]:DatenVorsprung/LRTNG.git
cd LRTNG
cmake ./
make all
To specify a model, e.g. brusselator, you need to create the files brusselator_fdyn.txt and brusselator_init.txt in
the folder Benchmarks. In the
_fdyn.txt file you define the nonlinear dynamical system function and in the _init.txt file,
you specify some model-specific arguments:
-
dimensiondimension of the nonlinear dynamical system -
ìnitial_radiusradius$\delta_0$ of the initial ball -
last_variable_is_time_t(boolean)If your benchmark has time as a variable (like our Dubins Car), put the time variable as the last one and set this argument to 1, otherwise 0 -
control_variables_without_radius_(last_ones_except_time_variable)If your benchmark has control variables or other variables whose value should be initialized by an exact value (without putting a ball around it), put them as the last ones (before time variable) and set this argument to the number of those variables. Example: Quadcopter -
initial_valuescenter$x_0$ of the initial ball
All models used in our paper are in the folder Benchmarks, so you can try them out before running LRT-NG with your own models. For example,
// cartpole_fdyn.txt (CartPole-v1 with a linear policy)
Function fdyn(dth, dx, th, x)
M = 1.0;
g = 9.81;
h = 0.0001;
l = 1.0;
m = 0.001;
f = -1.1 * M * g * th - dth;
fdth = 1.0 / (l*(M+m*sin(th)*sin(th))) * (f * cos(th) - m*l*dth*dth*cos(th)*sin(th) + (m+M)*g*sin(th));
fdx = 1.0 / (M+m*sin(th)*sin(th)) * ( f + m*sin(th) * (-l * dth*dth + g*cos(th)) );
fx = dx;
fth = dth;
return (fdth, fdx, fth, fx);
end
// cartpole_init.txt (CartPole-v1 with a linear policy)
Cartpole
********************
dimension:
4
********************
initial_radius:
0.01
********************
last_variable_is_time_t(boolean):
0
********************
control_variables_without_radius_(last_ones_except_time_variable):
0
********************
initial_values:
dth,dx,th,x
0
0
0.001
0The entry point of LRT-NG is defined in the file main.cpp, which accepts several arguments to specify properties of the reachtube. The most notable arguments are:
| argument | description | default value |
|---|---|---|
--help |
produce help message | false |
--benchmark |
choose benchmark | bruss |
--time_horizon |
set time horizon T | 10 |
--time_step |
set time step dt | 0.01 |
--integration_order |
Runge-Kutta integration order (1, 2 or 4) | 1 |
--compute_volume |
compute and output volume of Reachtube (bool) | 1 |
--print_output |
print output to console (bool) | 1 |
You can use the run.sh script to define the parameters and use it to run several models or distinct settings consecutively:
./run.sh
There are several _output.txt files in saved_outputs. You can plot the reachtube by running plot_LRTNG.py using the _ellipse_output.txt and _circle_output.txt files.
cd plot_results
python3 -m venv venv # Optional
source venv/bin/activate # Optional
python3 -m pip install -r requirements.txtThe entry point of plotting is defined in the file plot_LRTNG.py, which accepts several arguments to specify properties. The required arguments are:
| argument | description |
|---|---|
--dim |
dimension of the nonlinear dynamical system |
--time_horizon |
time horizon T of plot |
--time_step |
used time step dt |
--ellipse_file |
name of the ellipse file inside --output_directory |
--circle_file |
name of the circle file inside --output_directory |
The last lines of the terminal output after running the code shows how to plot that specific Reachtube. For example:
cd plot_results
source venv/bin/activate
python plot_LRTNG.py --dim 2 --time_horizon 10 --time_step 0.01 --circle_file 1636844712_circle_output.txt --ellipse_file 1636844712_ellipse_output.txt
deactivate
cd ..The number of successful iterations made and the volume of the reach-sets is saved in _volume_output.txt file inside saved_outputs as well as part of the terminal output:
Volume after 1001 iterations:
Average Volume: 1.1138057e-07
Initial Volume: 3.0842514e-09
Final Volume: 4.0225955e-07
@INPROCEEDINGS{gruenbacher2020lagrangian,
author={Gruenbacher, Sophie and Cyranka, Jacek and Lechner, Mathias and Islam, Md Ariful and Smolka, Scott A. and Grosu, Radu},
booktitle={CDC},
title={Lagrangian Reachtubes: The Next Generation},
year={2020},
pages={1556-1563},
}