Machine learning - learn to add

learn_to_add.ts

import * as tf from '@tensorflow/tfjs'

/*
 * Experiment with
 *  - multiplication, combined (ie x1+3*x2). May have to train 500 epochs to get satisfactory results.
 */

const randomInt = (max: number) =>
  Math.floor(Math.random() * Math.floor(max))

const randomInts = (max: number, n: number): number[] =>
  new Array(n).fill(0).map(_ => randomInt(max))

const unknown_function = (x1: number, x2: number) =>
  x1 + x2

const generateSamples = (max: number, n: number) => {
    const x1s = randomInts(max, n)
    const x2s = randomInts(max, n)
    return {
        xs: x1s.map((_, i) => [x1s[i], x2s[i]]),
        ys: x1s.map((_, i) => unknown_function(x1s[i], x2s[i])),
    }
}

type Data = number[]
type Observation = number
type Prediction = Observation

async function learn(xs: Data[], ys: Observation[]) {
    const model = tf.sequential({
        layers: [
            tf.layers.dense({ units: 16, activation: 'relu6', inputShape: [2] }),
            tf.layers.dense({ units: 16, activation: 'relu6' }),
            tf.layers.dense({ units: 1}),
        ],
    })

    model.compile({ optimizer: 'adam', loss: 'meanSquaredError' })

    await model.fit(tf.tensor2d(xs, [xs.length, 2]), tf.tensor2d(ys, [ys.length, 1]), { epochs: 50 })
    return (x: Data): Prediction => {
        // use xs and ys to predict outcome of x
        const tx = tf.tensor2d(x, [1,2])
        const f = model.predict(tx)
        return model.predict(tx).dataSync()[0]
    }
}

const { xs, ys } = generateSamples(10, 500)

console.log('learning...')
learn(xs, ys).then(predict => {
  console.log('finished learning.')
  console.log('#### 1+3 = ', predict([1, 3]))
  console.log('#### 5+2 = ', predict([5, 2]))
  console.log('#### 6+7 = ', predict([6, 7]))
  console.log('#### 9+9 = ', predict([9, 9]))
})