NeuralOC

Pytorch implementation of our neural network approach for solving optimal control problems

Additional videos at https://imgur.com/a/eWr6sUb

Associated Publications

A Neural Network Approach for High-Dimensional Optimal Control

Full-Paper: https://ieeexplore.ieee.org/abstract/document/9786046

Pre-Print: https://arxiv.org/abs/2104.03270

Please cite as

@article{onken2022neural,
  title     = {A Neural Network Approach for High-Dimensional Optimal Control Applied to Multiagent Path Finding},
  author    = {Derek Onken and Levon Nurbekyan and Xingjian Li and Samy Wu Fung and Stanley Osher and Lars Ruthotto},
  journal   = {IEEE Transactions on Control Systems Technology},
  year      = {2022},
  publisher = {IEEE},
  doi       = {10.1109/TCST.2022.3172872}
}

A Neural Network Approach Applied to Multi-Agent Optimal Control

Conference Paper: https://ieeexplore.ieee.org/abstract/document/9655103

Pre-Print: https://arxiv.org/abs/2011.04757

Please cite as

@inproceedings{onken2020neural,
    title     = {A Neural Network Approach Applied to Multi-Agent Optimal Control},
    author    = {Derek Onken and Levon Nurbekyan and Xingjian Li and Samy Wu Fung and Stanley Osher and Lars Ruthotto},
    booktitle = {European Control Conference (ECC)},
    pages     = {1036--1041},
    year      = {2021},
    doi       = {10.23919/ECC54610.2021.9655103}
}

Set-up

Install all the requirements (designed for python 3.6.9):

pip install -r requirements.txt

Try a simple command

python trainOC.py

More elaborate instructions available at detailedSetup.md.

Experiments

Neural Network Examples

Pretrained neural network examples

Corridor Experiment

python evalOC.py --nt 50 --resume experiments/oc/pretrained/softcorridor_nn_checkpt.pth
  
  # additional flag to produce the shocks for the corridor problem
python evalOC.py --nt 50 --resume experiments/oc/pretrained/softcorridor_nn_checkpt.pth --do_shock
  
  # additional flag to produce video
python evalOC.py --nt 50 --resume experiments/oc/pretrained/softcorridor_nn_checkpt.pth --make_vid

2-agent Swap Experiment

python evalOC.py --nt 50 --resume experiments/oc/pretrained/swap2_nn_checkpt.pth --make_vid

12-agent Swap Experiment

python evalOC.py --nt 50 --resume experiments/oc/pretrained/swap12_nn_checkpt.pth --make_vid

Swarm Experiment

python evalOC.py --nt 80 --resume experiments/oc/pretrained/swarm50_nn_checkpt.pth --make_vid

Quadcopter Experiment

python evalOC.py --nt 50 --resume experiments/oc/pretrained/singlequad_nn_checkpt.pth --make_vid

Train your own neural networks

Corridor Experiment

python trainOC.py --data softcorridor --lr_freq 600 --lr 0.01 --niters 1800 --lr_decay 0.1 --prec single --nt 20 --nt_val 32 --nTh 2 --var0 1.0 --log_freq 10 --val_freq 25 --viz_freq 100 --m 32 --n_train 1024 --sample_freq 100 --alph 100.0,10000.0,300.0,0.02,0.02,0.02

2-agent Swap Experiment

python trainOC.py --data swap2 --lr_freq 1000 --lr 0.01 --niters 4000 --lr_decay 0.1 --prec single --nt 20 --nt_val 42 --nTh 2 --var0 1.0 --log_freq 10 --val_freq 50 --viz_freq 100 --m 16 --n_train 1024 --sample_freq 50  --approach ocflow  --gpu 0 --alph 300.0,1000000.0,100000.0,1,1,3

12-agent Swap Experiment

python trainOC.py --data swap12 --lr_freq 800 --lr 0.01 --niters 2400 --lr_decay 0.1 --prec single --nt 20 --nt_val 32 --nTh 2 --var0 1.0 --val_freq 50 --viz_freq 100 --m 32 --n_train 2048 --sample_freq 100  --approach ocflow --gpu 0 --alph 300.0,0.0,100000.0,5.0,2.0,5.0

Swarm Experiment

python trainOC.py --data swarm50 --lr_freq 1200 --lr 0.001 --niters 6000 --lr_decay 0.5 --prec single --nt 26 --nt_val 50 --nTh 2 --var0 0.1 --val_freq 25 --viz_freq 500 --m 512 --n_train 1024 --sample_freq 25 --approach ocflow  --gpu 0 --alph 1800.0,10000000.0,25000.0,2,1,3 --new_alph 800,900.0,10000000.0,25000.0,0.0,0.0,0.0

Quacopter Experiment

python3 trainOC.py --data singlequad --lr_freq 1200 --lr 0.01 --niters 6000 --lr_decay 0.5 --prec single --nt 26 --nt_val 50 --nTh 2 --var0 0.1 --val_freq 50 --viz_freq 100 --m 128 --n_train 1024  --sample_freq 25  --alph 5000.0,0.0,0.0,0.1,0.0,0.0

Baseline Examples

Pretrained baseline examples

Corridor Baseline presented in publication

python baseline2D.py --data softcorridor --nt 50 --alph 100.0,10000.0,300.0 --resume experiments/oc/pretrained/softcorridor_baseline_checkpt.pth

Baseline and Neural Network comparisons presented in publications

python compareCorridor.py

python compareQuad.py

Train your own baseline examples

python baseline2D.py --data softcorridor --nt 50 --alph 100.0,10000.0,300.0

python baselineQuad.py --data singlequad --nt 50 --alph 5000.0,0.0,0.0

Compare Real-Time Timings of NN and optimizer-invariant baseline

We encapsulate all the commands into a bash script and restrict the running to a single CPU core:

taskset -c 0 ./scriptTimeDeployment.sh

Acknowledgments

This material is in part based upon work supported by the US National Science Foundation Grant DMS-1751636, the US AFOSR Grants 20RT0237 and FA9550-18-1-0167, AFOSR MURI FA9550-18-1-050, and ONR Grant No. N00014-18-1- 2527. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the funding agencies.

donken/NeuralOC