JNN

Java Library for neural networks with evolutionary training

there is a java and c# version

• java fully tested (needs "Apache common math" for CPU and "JCuda" for Nvidia GPU)
• C# version in beta because i didn't test it (and is based on a really old Java version of this lib)

Example

The Java version has an Example, where the black rectangle needs to eat the green food and avoid the red enemies

api

• Setup
• Training
• save NN
• info

Setup

.setup(int[] layout, int AgentCount, int Generationsteps, int MutationRate, boolean cuda);

• layout describes the NN. for example:

   {2,2,2} means NN has 2 inputs
  2 hidden neurons in one layer
  2 outputs
  

• Generationsteps is the Amount of times it makes a generation before actually making children, so no AI can get to the next generation by luck

• If cuda is true the NN will run on your Nvidia GPU. Note: for small NNs CPU is recommended

Training

• set Input

   .setInput(int Input, double Value, int Agent);
  

• run Agent

   .runAgent(int Agent);
  

• run Agent with multithreading

   .runAllAgents();
  

• get Output (returns double)

   .getOutput(int Output, int Agent)
  

• set Fitness

   .setFitness(float Fitness, int Agent);
  

• next Generationstep

   .nextGenerationstep();
  

saveNN

• save NN

   .saveNN(int Agent);
  

• save best NN

   .saveNN(JNN.getAgentIDPerFitnessList(0));
  

info

• get Generation (returns int)

   .getGeneration()
  

• get Generationstep (returns int)

   .getGenerationstep()
  

• get Fitness(returns float)

   .getAgentFitnessPerFitnessList(int Number)
  

• get Agent ID orderd with Fitness (returns int)

   .getAgentIDPerFitnessList(int Number)