nn.js

let trainExp = 7;
let datasetSize = 2000;
let printValidateEvery = 100;
let epoches = 1000;
let validateRatio = 0.2;
let accuracyThreshold = 0.99999;
let shouldPassAccuracyThresholdAtLeastEpoch = 300;

let network;

function evalNN(network, data) {
  let [h1r, h2r, h3r, h4r, xh1, xh2, xh3, xh4, yh1, yh2, yh3, yh4] = network;
  let [trainX, trainY, trainR] = data;
  let [h1Out, h2Out, h3Out, h4Out] = [
    (trainX * xh1) + (trainY * yh1),
    (trainX * xh2) + (trainY * yh2),
    (trainX * xh3) + (trainY * yh3),
    (trainX * xh4) + (trainY * yh4)
  ];
  let rOut = (h1r * h1Out) + (h2r * h2Out) + (h3r * h3Out) + (h4r * h4Out);
  return [h1Out, h2Out, h3Out, h4Out, rOut];
}

function toRange(x, y) {
  let exp = Math.ceil(Math.log10(x + y));
  let entry = [x / (10 ** exp), y / (10 ** exp), (x + y) / (10 ** exp)];
  return [exp, entry];
} 

function trainNN() {
  // Generate dataset
  let trainDataset = [];
  let validateDataset = [];
  for (let i = 0; i < datasetSize; i++) {
    let x = Math.random() * (10 ** trainExp);
    let y = Math.random() * (10 ** trainExp);
    let entry = toRange(x, y)[1];
    if (i < datasetSize * validateRatio) {
      validateDataset.push(entry);
    } else {
      trainDataset.push(entry);
    }
  }

  // init weights
  let [h1r, h2r, h3r, h4r, xh1, xh2, xh3, xh4, yh1, yh2, yh3, yh4] = [
    Math.random(), Math.random(), Math.random(), Math.random(),
    Math.random(), Math.random(), Math.random(), Math.random(),
    Math.random(), Math.random(), Math.random(), Math.random()
  ];

  // las-vegas training
  let reinitIfNeeded = (v) => v < -10 ? Math.random() : (v > 10 ? Math.random() : v);
  for (let i = 1; i <= epoches; i++) {
    for (let data of trainDataset) {
      let [h1Out, h2Out, h3Out, h4Out, rOut] = evalNN([h1r, h2r, h3r, h4r, xh1, xh2, xh3, xh4, yh1, yh2, yh3, yh4], data);
      let trainR = data[2];
      let eR = trainR - rOut;
      let [eH1, eH2, eH3, eH4] = [eR * h1r, eR * h2r, eR * h3r, eR * h4r];
      [h1r, h2r, h3r, h4r, xh1, xh2, xh3, xh4, yh1, yh2, yh3, yh4] = [
        reinitIfNeeded((eR * rOut) + h1r), reinitIfNeeded((eR * rOut) + h2r), reinitIfNeeded((eR * rOut) + h3r), reinitIfNeeded((eR * rOut) + h4r),
        reinitIfNeeded((eH1 * h1Out) + xh1), reinitIfNeeded((eH2 * h2Out) + xh2), reinitIfNeeded((eH3 * h3Out) + xh3), reinitIfNeeded((eH4 * h4Out) + xh4),
        reinitIfNeeded((eH1 * h1Out) + yh1), reinitIfNeeded((eH2 * h2Out) + yh2), reinitIfNeeded((eH3 * h3Out) + yh3), reinitIfNeeded((eH4 * h4Out) + yh4),
      ];
    }
    // validate
    if (i % printValidateEvery === 0) {
      let diffs = [];
      for (let data of validateDataset) {
        let [h1Out, h2Out, h3Out, h4Out, rOut] = evalNN([h1r, h2r, h3r, h4r, xh1, xh2, xh3, xh4, yh1, yh2, yh3, yh4], data);
        let trainR = data[2];
        diffs.push(Math.abs(trainR - rOut));
      }
      let sum = diffs.reduce((a, b) => a + b);
      let avg = sum / diffs.length;
      let accuracy = 1 - Math.abs(avg);
      console.log(`Epoch ${i}, Accuracy: ${1 - Math.abs(avg)}`);
      if (i >= shouldPassAccuracyThresholdAtLeastEpoch && accuracy < accuracyThreshold) {
        return trainNN(); // who cares
      }
    }
  }
  return [h1r, h2r, h3r, h4r, xh1, xh2, xh3, xh4, yh1, yh2, yh3, yh4];
}

function add(x, y) {
  if (!network) {
    network =  trainNN();
  }
  let [exp, range] = toRange(x, y);
  let r = evalNN(network, range)[4];
  return Math.round(r * (10 ** exp));
}

console.log(add(1, 2));
console.log(add(10, 10));
console.log(add(111, 222));
console.log(add(999, 999));
console.log(add(123456789, 987654321));
console.log(network);