ml-js v0.0.10

ml-js

Machine Learning library for Node.js

Status : Under development

Installation

ml-js depends on FANN (Fast Artificial Neural Network Library) witch is a free, open source and high performence neural network library.

To build great app with it :

• Make sure you glib2 is installed : sudo apt-get install glib2.0
• make sure pkg-config is installed : sudo apt-get install pkg-config
• make sure cmake is installed : sudo apt-get install cmake
• Install FANN :
  • download here
  • unzip
  • goto to FANN directory
  • run cmake . and sudo make install
  • run sudo ldconfig

Finally, you should be able to install all npm dependancies with npm install ml-js --save

Supported ML techniques

ml-js currently supports :

• Supervised learning :
  • fann - Neural Networks provided by node-fann addon
  • svm - Support vector machine provided by node-svm addon
• Reinforcement learning :
  • QLearning, model-free reinforcement learning technique
• Exploration policies
• BoltzmannExploration

Getting started

QLearning Example

Q-learning is a model-free reinforcement learning technique. Specifically, Q-learning can be used to find an optimal action-selection policy for any given MDP.

ml = require 'ml-js'

myprocess = new SomeProcess

qValues = new ml.CSDAQValues nb_features, nb_actions

options = {
  learning_rate: 0.1
  discount_factor: 0.9
  exploration_policy: new ml.BoltzmannExploration 0.2 # temperature
}
agent = new ml.QLearningAgent qValues, options

myprocess.on 'do_something', (currentState)->
  next_action = agent.getAction currentState
  myprocess.do next_action
  
myprocess.on 'feedback_received', (initState, action, newState, reward)->
    agent.learn initState, action, newState, reward

Availables QValues managers :

• CSDAQValues for Continuous States and Discrete Actions QValues. It uses neural networks to perform learning on continuous states.