This is my C++ implementation of hypercube based extension of NEAT (NeuroEvolution of Augmenting Topologies) algorithm originally invented by Dr. Kenneth O. Stanley. Implementation developed as my ability and understanding of the topic allowed.
For the details about the algorithm and other implementations please visit the official users page at http://eplex.cs.ucf.edu/hyperNEATpage/HyperNEAT.html
- HyperNEAT-LEO variation. CPPN takes 4 inputs and bias node (two 2D coordinates + 1.0) and outputs 4 values: weight expression, weight value, bias expression, bias value.
- Disqualifying solutions returning NaN value for fitness.
- In-species champions.
- In-species selection odds based on fitness.
- Permanent innovation archive.
- Self-adapting mutation step for connection weights.
- Skipping solutions without connections.
- Species hint to optimise species assignments.
- Two benchmarks.
- g++ with C++14 support.
- Ubuntu.
If the build does not work, it should be relatively easy to set up a project using source files as it has no dependencies other than STL.
Run make command in your shell to build the package. Parameters
located in settings folder can be adjusted to achieve best (or
worst) evaluation counts.
By default the application will try to construct a neural network
to correctly classify examples from the Iris dataset. To run XOR
benchmark, uncomment the line in main() located in
source/main.cpp.
- XOR function: simple nonlinear function meant for preliminary verification and debugging. Performs 10 runs and computes the average evaluations per solution.
- Iris dataset: simple classification challenge for further verification of the implementation. Dataset source can be found here. Output best viewed on 80 x 24 terminal display.
Method descriptions are in .h files.