Duplex Optimization

Duplex's website

Duplex is a high-performance global optimization algorithm for nonconvex, nonlinear, and functional optimization problems. The following core capabilities are included:

• Duplex can find global optimum of nonlinear non-convex functions.
• Duplex implements variety of gradient-descent based optimizations internally, such as Momentum, AdaDelta, Adamax. Duplex can also optimize when the gradient information are not available.
• Duplex supports unsupervised learning algoriths for clustering.
• If the gradient information are not available (for circuit optimization), duplex uses reinforcement learning to predict the landscape of energy function.
• Duplex supports Synopsys HSPICE for solving nonlinear systems.

The latest release and a complete manual may be found at the Duplex home page: http://adel.ac/duplex

Dependencies

Duplex requires the following dependencies to be installed:

• Eigen 3.3.0 linear algebra library. Install eigen using brew install eigen
• Boost 1.51 Duplex requires both headers and compiled binaries. Install using brew install boost
• Config4cpp configuration library. The binaries for the config4cpp for macOS are already shipped with Duplex (/submodules/config4cpp/lib).
• Pangolin for plotting and managing display. The Pangolin will replace the gnuPlot. Duplex will install Pangolin as a submodule in /submodules/Pangolin.
• GnuPlot for drawing plots (Optional). Install gnuplot using brew install gnuplot

Build

Duplex uses CMake as a build system.

git clone git://github.com/ahmadyan/Duplex
cd Duplex
mkdir build	
cd build
cmake ..
make

Use cmake .. -GXcode to generate the xcode project file (or other generators). Few examples are provided in /bin directory. Few MATLAB tests are available in the /test directory.

Usage

• ./duplex --config example.cfg
• Currently the duplex binaries and the configurations has to be in the same directory, so please output the binary in the bin folder.
• Example configuration files are available in the /bin directory
• If using hspice as a numerical simulator, User should set an environment variable DUPLEX_SIM_ID to an integer value. If you don't want to set the environment variable, run Duplex from the python wrapper (test/duplex.py) in test directory. Python wrapper requires Python 3+.

Misc

Duplex uses llvm coding style. Any PR will be formatted using clang-format.