gradientDescent.js

/* ----------------------------------------- *
        Learning Machin Learning
* ----------------------------------------- */
// gradientDescent for the square error function and a Theta of Two parameter (a, b)

// Theta :: [Number, Number]

// sum :: (a -> Int) -> List a -> Int
const sum = (transformer, list) =>
  list.reduce((acc, item) => acc + transformer(item), 0)


// getNextTheta :: Number -> List Data -> Theta
const getNextTheta = (step, dataSet) => ([a, b]) => [
  a - step / dataSet.length * sum(([x, y]) => a + b * x - y, dataSet),
  b - step / dataSet.length * sum(([x, y]) => (a + b * x - y) * x, dataSet),
]


// isConverging :: Theta -> Theta -> Bool
const isConverging = ([a1, b1], [a2, b2]) =>
  Math.abs(a1 - a2) < 0.01 && Math.abs(b1 - b2) < 0.01


// repeatUntilConverge :: (Theta -> Theta) -> Theta -> Theta
const repeatUntilConverge = (f, theta) => {
  const nextTheta = f(theta)
  return isConverging(nextTheta, theta)
    ? nextTheta
    : repeatUntilConverge(f, nextTheta)
}

// gradientDescent :: { step :: Number, dataSet :: List Data } -> Theta
const gradientDescent = ({ step, dataSet }) =>
  repeatUntilConverge(getNextTheta(step, dataSet), [0, 0])



export default gradientDescent




// linearRegression.js

import gradientDescent from './gradientDescent'

// linearRegression :: List Data -> (Number -> Number)
export default function linearRegression(dataSet) {

  const [a, b] = gradientDescent({ step: .0000001, dataSet })
  console.log(a, b)

  return x => a + b * x
}