cpp_starter

Various optimization algorithms

Dependencies

• a modern C++ compiler (i.e. one that supports C++17 features, ex. gcc 7+)
• cmake 3.13 or newer
• Eigen3 (installable via apt, yum, or homebrew)

Building

From the project's root directory,

mkdir build
cd build
cmake ..
make

Tips:

• you can build using multiple threads at once using e.g. make -j4
• you can specify a specific compiler using cmake .. -DCMAKE_CXX_COMPILER=g++-8
• for debugging symbols use cmake .. -DCMAKE_BUILD_TYPE=Debug

TODO

save data to files and run python with a make command run with single size parameter to see what happens

For the first thing... an EVO experiment needs pop size, learning rate, n iterations, print stride (common) and a starting distribution (depends on type of distribution) separate command line tools for each experiment?

objective for optimization gives input type and gradient type, can eval both (apart and at once)

• should objective give actual input type, or const ref? For now it'll always be const &, so I'll add those elsewhere (need to be able to make the literal thing in some contexts)
• gradient type can be multiplied by scalar; input type can be += with gradient

expected value estimator?

• stores n_samples
• takes objective and distribution (objective only needs to eval points)
• input type is distribution::params
• gradient type is distribution::score (hm should I call this gradient?)
• can estimate value at params
• can estimate gradient wrt params at params

then I'll be able to change just the distribution - expected value

how to handle data logging?