Neatzsche is a implementation of Kenneth Stanley's NEAT algorithm. The implementation was started as a part of Bjørn Magnus Mathisens master thesis in 2007. Part of it was written at NTNU in Trondheim and parts of it was written while visiting Ken's research group at UCF.

The implementation is done in C++ and supports MPI for running in paralell; The fitness evaluation is distributed, while the genetic computations are done at the central server. The user can instruct the code to give the "master" node only a given percentage of the fitness-evaluation workload of the other nodes. This way one can guarantee that the master node is finished with it's fitness evaluations before the results come in from the other nodes in the MPI network.

The code has been used in two master thesis:

• My own: implementing the MPI-distributed NEAT in c++ and finding it's usefullness in evaluating how a "roving eye" can play Go.
• Master thesis by Anders Storsveen: Where he uses my NEAT. Link: implementation to show that NEATS usage as a CPPN (compositional pattern producing networks) can create an eye through evolution. Link:

The code was created for research, and as such has poor interface design. It has many command line arguments some of which are nested into substrings depending on previous arguments - it is all however documentet when running the binary.

The thesis describing the first implementation and experiments using it to play go can be found here. Anders Storsveen's master thesis describing the usage of this software as a CPPN to create eye can be found here

usage: (all the node numbers are for the initial generation) 
neatzsche [seed (0 for new)] [settings file] [pop options] [selector option] [phenotype eval sofile] [phenotype eval options] [stopcondition] [mpi run(0/1)] [speciationgraph (0/1)]
where:
	[pop options] = "spawn [input nodes] [output nodes] [hidden nodes] [link probability] [population size]"
	[pop options2] = "genesis [genome-file] [population size] [initial elitism]"
	[selector options] = "fp [number of top species getting elitism]"
	[selector options2] = "rank [s] [number of top species getting elitism]"
	[selector options3] = "sigma [number of top species getting elitism]"
	[stop condition1] = "count [generations] [runs]"
	[stop condition2] = "key"

The fitness evaluation is implemented as a plugin system, so you only implement a given "api" so that the NEAT algorithm can call you fitness evalutation.

This plugin structure follows a pattern:

#include "eval.h"
#include "iface.h"
extern "C" {
  FitnessEvaluator *maker(char * str)
  {
    FitnessEvaluator * ret;
    vector<string> * sv = split(str," ");
    if(sv->at(0).find("dataset")!=string::npos) {
      if(sv->size()!=4){
      cerr << "wrong arguments to dataset eval should be: \"dataset <datasetfile> [classification at start[0/1]] [testration]\"" << endl;
	exit(1);
      }
      bool classAtStart = (atoi(sv->at(2).c_str()) == 1) ? true : false;
      double testratio = atof(sv->at(3).c_str());
      DataSet * ds = new DataSet(classAtStart,sv->at(1),testratio);
      ret = new DatasetEvaluator(ds);
      return ret;
    }else
      cerr << "wrong arguments to dataset eval should be: \"dataset <datasetfile> [classification at start[0/1]] [testration]\"" << endl;
      return NULL;
  }
}

more examples of fitness evaluation plugins can be found in the code repository under expeeriments