a test of detecting chart type using neural net

Rakefile

@@ -78,7 +78,7 @@ namespace :test do
   MOCHA      = "./node_modules/.bin/mocha"
   COVERALLS  = "./node_modules/.bin/coveralls"
 
-  OPTS = "--compilers coffee:coffee-script/register --ui exports --bail"
+  OPTS = "--compilers coffee:coffee-script/register --ui exports --timeout 5000 --bail"
 
   desc "Run code coverage, mocha with Blanket.js"
   task :coverage do

package.json

@@ -21,6 +21,8 @@
   },
   "dependencies": {
     "async": "~0.9.0",
+    "brain": "^0.7.0",
+    "chance": "^0.6.1",
     "d3": "~3.4.13",
     "d3-tip": "Caged/d3-tip",
     "director": "~1.2.3",

test/neural.coffee

@@ -0,0 +1,59 @@
+{ assert } = require 'chai'
+Chance     = require 'chance'
+brain = require 'brain'
+_     = require 'lodash'
+d3    = require 'd3'
+
+chance = new Chance()
+
+module.exports =
+
+  'neural net - detect sine chart type': (done) ->
+    fns =
+      # A "rough" sine wave.
+      sine: (x) -> Math.sin chance.floating 'min': Math.max(x - 1, 0), 'max': x + 1
+      # A linear function.
+      linear: (x) -> ( chance.floating 'min': Math.max(x - 1, 0), 'max': x + 1 ) + 2
+
+    # Generate sample data points.
+    samples = 10
+    data = _.map [ 0 ... 10 ], (i) ->
+      fn = fns[ [ 'linear', 'sine' ][ sine = +(i < (samples / 2)) ] ]
+      points = ( fn(j) for j in [ 0 ... 100 ] )
+      { points, sine }
+
+    # Scale the data to just 6 points.
+    for chart in data then do (chart) ->
+      step = Math.round (len = chart.points.length) / 5
+      chart.points = ( chart.points[ Math.min(i, len - 1) ] for i in _.range(0, len + step, step) )
+
+    # Scale the values for each chart.
+    for chart in data then do (chart) ->
+      scale = d3.scale.linear().range([ 0, 1 ]).domain([ d3.min(chart.points), d3.max(chart.points) ])
+      chart.points = _.map chart.points, scale
+
+    # Split the dataset.
+    training = data[ 1...9 ]
+    testing  = [ data[0] ].concat data[9]
+
+    net = new brain.NeuralNetwork()
+
+    keys = [ 0, 0.2, 0.4, 0.6, 0.8, 1 ]
+    input = for chart in training then do (chart) ->
+      'input': _.zipObject keys, chart.points
+      'output':
+        'sine': chart.sine
+    
+    net.train input,
+      'errorThresh':  0.005 # error threshold to reach
+      'iterations':   5e4   # maximum training iterations
+      'log':          no    # console.log() progress periodically
+      'logPeriod':    1e3   # number of iterations between logging
+      'learningRate': 0.2   # learning rate
+
+    # Now test it.
+    for sample in testing
+      out = net.run _.zipObject keys, sample.points
+      assert sample.sine is Math.round(out.sine), "Network says #{out.sine} but sample says #{sample.sine}"
+
+    do done