NOA: Nonlinear Optimisation Algorithms

We aim to make it easier to integrate optimisation with on-line learning, bayesian computation and large simulation frameworks through dedicated differentiable algorithms. Our solution is suitable for both research and applications in performance demanding systems such as encountered in streaming analytics, game development and high frequency trading.

Installation

Currently, we support only GNU, and CUDA for GPU (check WSL for Windows, and LIMA for macOS). A toolchain fully supporting C++17 is required. NOA is a header-only library, so you can directly drop the src/noa folder into your project.

:warning: However, beyond C++17 source code, the project contains:

• CUDA sources in noa/kernels.cuh
• C++ third-party sources in noa/kernels.hh

which may require separable compilation.

Note 1

src/noa/kernels.hh could be compiled separately via test/kernels.cc, but it is also possible to include it directly into your program if you want to keep everything inside of the single translation unit.

Note 2: PUMAS

PUMAS code under src/noa/3rdparty/_pumas was syntactically adapted to compile under g++ with -fpermissive. To avoid setting this parameter everywhere, noa/kernels.hh doesn't include PUMAS by defaut. To include PUMAS, define NOA_3RDPARTY_PUMAS before including noa/kernels.hh, and compile the corresponding file with an -fpermissive flag.

CMake project

The core of the library depends on LibTorch (which is also distributed via pip and conda).

You can specify whether torch was compiled with C++11 ABI setting TORCH_USE_CXX11_ABI to 0 or 1. If we manage to fetch the value of _GLIBCXX_USE_CXX11_ABI during configuration, your choice will be overwritten. By default, this variable is 0.

For additional configuration needed by some applications please refer to the documentation below.

We encourage you to work with conda. If your system supports CUDA, the environment env.yml contains all the required libraries:

$ conda env create -f env.yml
$ conda activate noa

For a CPU only installation please use env-cpu.yml instead.

Build tests & install the library (to turn testing off add -DBUILD_NOA_TESTS=OFF):

$ mkdir -p build && cd build
$ cmake .. -GNinja -DCMAKE_INSTALL_PREFIX=$CONDA_PREFIX
$ cmake --build . --target install
$ ctest -V

To build benchmarks specify -DBUILD_NOA_BENCHMARKS=ON. To enable parallel execution for some algorithms you should link against OpenMP. To build CUDA tests add -DBUILD_NOA_CUDA=ON and the GPU architecture of your choice, e.g. -DCMAKE_CUDA_ARCHITECTURES=75.

To build documentation you need to have doxygen installed and specify -DBUILD_DOCS=ON.

Finally, once NOA is installed, you can link against it in your own CMakeLists.txt file. Make sure to add LibTorch as well:

cmake_minimum_required(VERSION 3.12)
set(CMAKE_CXX_STANDARD 17)
find_package(Torch REQUIRED)
find_package(NOA CONFIG REQUIRED)
target_link_libraries(your_target torch NOA::NOA)
target_compile_options(your_target PRIVATE -Wall -Wextra -Wpedantic -O3)

KMath wrapper

NOA is exposed within the kotlin library KMath as a dedicated module kmath-noa. To build the JNI wrapper you need to add -DBUILD_JNOA=ON setting -DJAVA_HOME=<...> if necessary. This will produce the shared library jnoa to which you should point the java.library.path for the JVM to load it.

Applications

NOA offers several advanced applications for optimisation. Please refer to the documentation and usage examples for each component to find out more:

• GHMC focuses on Bayesian computation with the Geometric HMC algorithms dedicated to sampling from higher-dimensional probability distributions.
• PMS provides a framework for solving inverse problems in the passage of particles through matter simulations.
• CFD implements adjoint sensitivity models for a variety problems arising in computational fluid dynamics.
• QC is a differentiable quantum chemistry platform for materials design.
• QUANT a differentiable derivative pricing library.

Acknowledgments

When referring to this library in scientific work please cite

@article{grinis2022diffprog,
  title={Differentiable programming for particle physics simulations},
  author={Roland Grinis},
  journal={JETP Vol. 161 (2)},
  year={2022}
}

Contributors

• Sabina Abdiganieva researched various DFT implementations for QC.

• Sergey Barsegyan works on asymptotic implied volatility formulas for QUANT.

• Dmitry Bazanov developed the Heston volatility surface calibration routines in QUANT and works on its applications to real market data.

• Gregory Dushkin is the main developer for CFD, implementing differentiable programming algorithms for MHFEM schemes and integrating TNL into the library. He also works on muon simulations in PMS and intergrated pumas. He helped to implement the adjoint sensitivity algorithm for Kohn-Sham equations in QC.

• Anastasia Golovina wrapped LibTorch's optimizers into KMath.

• Eugen Kadilenko developed the adjoint sensitivity model for Kohn-Sham equations in QC.

• Andrey Lipsky works on applications to real market data for QUANT.

• Oleg Mosyagin implemented the XPU ray-tracing algorithm in PMS.

• Ivan Novikov contributes to the documentation in QUANT, implemented the PDE methods for the Black-Scholes model and works on margin models.

• Danila Ryazanov developed the XPU DCS computations for muons in PMS.

• Maksim Sosnin implemented the almost exact simulation scheme for the Heston Model in QUANT and works on integration with Deep Hedging.

• The JNI wrapper is being developed in collaboration with KMath contributors.

We welcome contributions to the project and would love to hear about any feature requests

License

NOA is licensed under the terms and conditions of the GNU General Public License (GPL) version 3 or - at your option - any later version. The GPL can be read online, or see the full LICENSE.

Please note that NOA license does NOT feature a template exception to the GNU General Public License. This means that you must publish any source code which uses any of the NOA header files if you want to redistribute your program to third parties. If this is unacceptable to you, please contact us for a commercial license.

For support or consultancy services contact GrinisRIT.

(c) 2023 GrinisRIT ltd. and contributors