cpp.md

Using the C++ library

The core of MultiNEAT is plain C++17 with no dependencies, and can be used without Python.

Building

cmake -S . -B build -DMULTINEAT_BUILD_PYTHON=OFF -DCMAKE_BUILD_TYPE=Release
cmake --build build --config Release
ctest --test-dir build -C Release        # runs the XOR example as a test
cmake --install build --prefix /usr/local  # optional: lib/libmultineat.a and include/MultiNEAT/*.h

In your own CMake project the simplest way is to add MultiNEAT as a subdirectory:

set(MULTINEAT_BUILD_PYTHON OFF CACHE BOOL "" FORCE)
set(MULTINEAT_BUILD_EXAMPLES OFF CACHE BOOL "" FORCE)
add_subdirectory(MultiNEAT)
target_link_libraries(my_experiment PRIVATE multineat)

Example

src/Main.cpp is a complete example. The essentials:

#include "Genome.h"
#include "Population.h"
#include "NeuralNetwork.h"
#include "Parameters.h"

using namespace NEAT;

Parameters params;
params.PopulationSize = 150;

GenomeInitStruct ints;
ints.NumInputs = 3;     // the last input is the bias
ints.NumOutputs = 1;
Genome seed(params, ints);

Population pop(seed, params, true, 1.0, /* RNG seed */ 0);

for (int generation = 0; generation < 300; generation++)
{
    for (auto &species : pop.m_Species)
    {
        for (auto &genome : species.m_Individuals)
        {
            NeuralNetwork net;
            genome.BuildPhenotype(net);
            std::vector<double> inputs = {1.0, 0.0, 1.0};
            net.Input(inputs);
            net.Activate();
            double fitness = /* ... */ net.Output()[0];
            genome.SetFitness(fitness);
            genome.SetEvaluated();
        }
    }
    pop.Epoch();
}

The classes and methods are the same as in Python (Population, Species, Genome, NeuralNetwork, Substrate, Parameters, RNG, Physics2D::Space ...); the Python names map directly to the C++ ones, and the members exposed as Python attributes are public C++ members with an m_ prefix (pop.Species is pop.m_Species, species.Individuals is species.m_Individuals).

C++ specific features:

• params.CustomConstraints - a bool (*)(Genome&) function returning true if a genome fails your topology constraints.
• Novelty search with PhenotypeBehavior subclasses: Population::InitPhenotypeBehaviorData() and Population::NoveltySearchTick().
• Serialization to strings: SerializeToString(obj) / DeserializeFromString(str, obj) from Serialization.h work for every class that can be pickled in Python.

Errors (bad files, invalid arguments) are reported with standard exceptions (std::runtime_error, std::invalid_argument). Compile with -DDEBUG to enable internal consistency checks (ASSERT).

Source layout

File
Genome.*	genomes, mutations, crossover, phenotype building, HyperNEAT
ESHyperNEAT.cpp	the ES-HyperNEAT phenotype builder
Genes.h, Traits.h	neuron/link genes and traits
Population.*, Species.*	the evolutionary algorithm
Innovation.*	the innovation database
NeuralNetwork.*	phenotypes, activation, learning
Substrate.*	HyperNEAT substrates
Parameters.*	all parameters, loading and saving
Random.*	the random number generator
Serialization.h	the archives used for pickling
Physics2D.*	the physics engine
PythonBindings.cpp	the pybind11 module (only compiled with PYTHON_BINDINGS)