janhohenheim / hippocrates

No longer maintained, actually usable implementation of NEAT

NO LONGER MAINTAINED

---

Branch	Linux	Windows
Master
Development

---

Visualization Tool by @Mafii
Experimental image recognition is currently being added at https://github.com/IDPA16/Hippocrates

Implementation of Kenneth Stanley and Risto Miikkulainen's NEAT (NeuroEvolution of Augmenting Topologies, http://nn.cs.utexas.edu/downloads/papers/stanley.ec02.pdf).

It focuses (in contrast to other implementations) on

• Speed - through modern and efficient C++17
• Clean Code - through constant ongoing refactoring and a deep care for aesthetics
• Usability - through being able to be used without much knowledge of Neural Networks
• Platform Independency - written on three different operating systems (Windows, Ubuntu, MacOS X) and the most used compilers (MSVC and GCC6), it is safe to say that it will work on multiple platforms, flawlessly.

Acknowledgements

• Serge A. Zaitsev (jsmn)

Dependencies

• ImageMagick 6.9.6-8

hippocrates's Issues

Output Neurons rarely have no incomming connections at all

Add CLang to Travis

Add the 'noexcept' identifier to applicable methods

Add functionality for semi-supervised training

Disable appveyor pull request builds

These builds are not required since appveyor builds each push anyway.

The CMakeLists are not configured to build a shared library

You did copy the CMakeLists.txt from JNF_NEAT_Examples

Serialization to Json uses newlines

This is not required and can be removed entirely. This will keep the json as small as possible.

See: cbd6926

Custom Weight Ranges are ignored

Perhaps the inputNeurons and outputNeurons Vectors are superfluous

Genome::AddRandomNeuron adds garbage

• gene.from can be higher than gene.to

NeuralNetwork::AddRandomConnection() adds backward connections

This is probably caused by NeuralNetwork::CategorizeNeuronsIntoLayers() not working for every genome

NeuralNetwork::AddRandomConnection() can connect to a Input Neuron

This is connected to the suboptimal design desicion to let outputs be in layer 0 and then count backwards

Reproduction should almost never be inter-species

It is possible for all childrens genes to be disabled

There is no Bias Neuron

It could be realized as just another inputNeuron.

Perturb weights using Gaussian Noise

Add support for recurrently selfconnected neurons

There constantly exists only exactly one Species

Add functionality for supervised training

move serialization and parsing to the stream operators

Add the ability to reset the networks memory while testing

Allow the training body to reset the memory of the network.

fitness logged in JSON is always 0

Parsing with `stoul` is not portable

std::size_t is not guaranteed to be an unsigned long.
Code like this will fail on some architectures.

Possible fix

Analyze the most demanding methods and add concurrency

Use std::mt19937 instead of rand()

Implement multithreading

It is quite important to gain performance on multi-core systems.

Add the new pendantic switch for Visual Studio

/permissive-
Please note the dash ( - ) !
Also add this for CNN

Add configuration for training without recursion

Run time Error

Not reproducible Runtime error Stackoverflow
function: NeuralNetwork::CategorizeNeuronBranchIntoLayers

Somewhere down the line all genomes weights get jumbled up to nonsensical values

Genome::ExtrapolateGeneCount is a faulty way to determine ranges for Gene insertion

This is part of the cause for #1 and #2

network does always output the same values

If asked for multiple outputs, the update method delivers the same output multiple times.

Implement image support through convolutionalism

Only allow merge commits

Implement a working Genome::AddRandomConnection()

The Incomming Connections of Neurons are pointing into garbage

This is (at least partly) caused by the vector of species not having a constant location. The references are getting destroyed during the process of inserting a new species and deleting empty ones

Build as DLL on VS

Return a more comprehensive error if a timeout occurs while testing

See #62

The produced Json for genomes has one depth too much:

Here is what the genomes in json look like:

,"genome":{"genome":[{"historicalMarking":0,"from":0,"to":3,"weight":0.252425....

But, it should be

,"genome": [{"historicalMarking":0,"from":0,"to":3,"weight":0.252425....

Why?

Because the json representing the neurons is:

{"neurons":[{"layer":0,"lastActionPotential":1.000000}

Speciate child at birth

Refactor In- and OutputNeuron accessors

apply DRY per method template with a lamba argument

Add load / save mechanisms

Linux builds fails because of std::experimental::filesystem

Unify indentation to tabs or spaces

Add functionality in supervised learning to check for overfitting

Recursive Connections are not permitted

Runtime Error

Error Runtime [Stackoverflow]
function NeuralNetwork::CategorizeNeuronBranchIntoLayers(Neuron& currNode)

GCC is having a SegFault

At this line since e01449c
Problem is probably older tho

Update Travis to GCC7

Genome::MutateGenes adds faulty new Genes

• The Method can connect Output Neurons to each other. Not Sure if this should be a functionality. Also I think there may be a off by one error.
• The Input Neurons can be connected