Stackelberg Game and Koopman Operator

This repo is for SG Koopman Learning project with an application in the guided navigation (trajectory planning).

Requirements

• Python 3.11

Running Scripts

1. Create a python virtual envrionment with Python 3.11 and source the virtual environment:

$ python3.11 -m venv <your-virtual-env-name>
$ source /path-to-venv/bin/activate

2. pip install the requirements:

$ pip install -r requirements.txt

3. In the project directory, first run the following script to generate data:

$ python exp_scripts/generate_data.py

4. In the project directory, run other scripts by commenting or uncommenting related functions:

$ python exp_scripts/training.py   # run training.py with train_kp_nn() as an example

5. (Optional) Create log directory and generate logs files by redicecting the output:

$ mkdir logs
$ mkdir logs/kp_nn         # example log directory
$ python exp_scripts/training.py > logs/kp_nn/log.txt

Comparison Algorithms

• kp_nn: Use Koopman operator to learn the follower's feedback dynamics.
• nn_fdynbr: Use neural network to learn the follower's feedback dynamics.
• dmd: Use DMD to learn the follower's feedback dynamics.
• nonlin_ocp: Use nonlinear optimal control (modle-based) to solve SE, serve as the baseline.

File Structure

• sg_koopman: Full algorithm implementations.
  • sg_koopman/common: Common classes and utilities.
• exp_scripts: Examples of calling modules and functions in sg_koopman.
  • Each script performs the function as the name suggests. E.g., training.py performs training for all algorithms.
  • The training (and RH planning) for each comparison algorithm is encapsulated by a function in each script. E.g., train_kp_nn() implements training for kp_nn algorithm.
  • Comment or uncomment functions to run each comparison algorithm.
• data: Store learning and planning results.
  • Each algorithm has an independent directory.

Coding Specifications

Individual agent's trajectory data is stored in a 3D numpy array with the format: D[i, k, :] = [x_k, u_k, x_kp1].

• D.shape[0] is the total number of trajectory.
• D.shape[1] is the length of each trajectory.

Note: x_kp1 in D[i,k,:] is the same as x_k in D[i,k+1,:]. We use the redundancy to align the data for fast access.

Interactive trajectory data is stored in a 3D numpy array with the format: D[i, k, :] = [xf_k, ufopt_k, xf_kp1, xl_k, ul_k, xl_kp1].

• The follower's control ufopt_k is optimal w.r.t. the leader's trajectory.
• Use the same data redundancy to align the data.