memlnaut-nisps/playground/js/nisps/iml.js
w1n5t0n 3db90b035a feat(playground): add spread param for sigmoid saturation control and fix randomise sync
Add ?spread=0-1 URL param that controls weight initialization scaling,
RL noise scaling per layer, noise cap, and weight decay to prevent
sigmoid output saturation. At spread=0 (original behavior) weights are
uniform [-1,1] and outputs polarise near 0/1. At spread=1 weights use
Xavier scaling (1/sqrt(fan_in)), noise is proportionally reduced, and
10% weight decay per thumbs-down prevents unbounded magnitude drift.

Also fix randomise to re-inject current joystick position and re-run
inference before routing outputs, eliminating the jump on first
joystick move after randomise.

Defaults: tame=1, spread=0.6 across all app variants.
2026-03-22 02:10:26 +02:00

217 lines
5.8 KiB
JavaScript

// NISPS IML - faithful port of nisps-core/include/nisps/iml.hpp + iml_impl.hpp
// Interactive Machine Learning interface
import { MLP } from './mlp.js';
import { Dataset } from './dataset.js';
export class IML {
/**
* @param {number} nInputs
* @param {number} nOutputs
* @param {number[]} hiddenLayers
* @param {number} maxIterations
* @param {number} learningRate
* @param {number} convergenceThreshold
*/
constructor(
nInputs,
nOutputs,
hiddenLayers = [10, 10, 14],
maxIterations = 1000,
learningRate = 1.0,
convergenceThreshold = 0.00001
) {
this.nInputs = nInputs;
this.nOutputs = nOutputs;
this.maxIterations = maxIterations;
this.learningRate = learningRate;
this.convergenceThreshold = convergenceThreshold;
// Build layer sizes: input+bias, hidden..., output
const BIAS = 1;
const layerSizes = [nInputs + BIAS, ...hiddenLayers, nOutputs];
// Activation functions: RELU for hidden, SIGMOID for output
const activationNames = [
...hiddenLayers.map(() => 'relu'),
'sigmoid',
];
this.dataset = new Dataset(100);
this.mlp = new MLP(layerSizes, activationNames);
this.inputState = new Array(nInputs).fill(0.5);
this.outputState = new Array(nOutputs).fill(0);
this.mode = 'inference';
this.performInference = true;
this.inputUpdated = true;
this.storedWeights = null;
this.weightsRandomised = false;
this.lastLoss = null;
this.bestLoss = null;
this.lossHistory = [];
this.totalTrainingIterations = 0;
this.logFn = null;
}
setLogger(fn) {
this.logFn = fn;
}
log(msg) {
if (this.logFn) this.logFn(msg);
}
setInput(index, value) {
if (index >= this.nInputs) return;
this.inputState[index] = Math.max(0, Math.min(1, value));
this.inputUpdated = true;
}
setInputs(values) {
for (let i = 0; i < values.length && i < this.nInputs; i++) {
this.inputState[i] = Math.max(0, Math.min(1, values[i]));
}
this.inputUpdated = true;
}
getOutputs() {
return this.outputState;
}
setOutput(index, value) {
if (index >= this.nOutputs) return;
this.outputState[index] = Math.max(0, Math.min(1, value));
}
setOutputs(values) {
for (let i = 0; i < values.length && i < this.nOutputs; i++) {
this.outputState[i] = Math.max(0, Math.min(1, values[i]));
}
}
process() {
if (!this.performInference || !this.inputUpdated) return;
// Add bias term
const inputWithBias = [...this.inputState, 1.0];
const { output } = this.mlp.getOutput(inputWithBias);
this.outputState = output;
this.inputUpdated = false;
}
getMode() {
return this.mode;
}
setMode(mode) {
if (mode === 'inference' && this.mode === 'training') {
this.train();
}
this.mode = mode;
}
// Two-step save example (hardware workflow)
saveExample() {
if (this.performInference) {
this.performInference = false;
this.log('Move to desired output position...');
return;
}
this.dataset.add(this.inputState, this.outputState);
this.performInference = true;
// Run inference
const inputWithBias = [...this.inputState, 1.0];
const { output } = this.mlp.getOutput(inputWithBias);
this.outputState = output;
this.log('Example saved.');
}
// Direct programmatic example addition
addExample(inputs, outputs) {
const inVec = inputs.slice(0, this.nInputs);
while (inVec.length < this.nInputs) inVec.push(0);
const outVec = outputs.slice(0, this.nOutputs);
while (outVec.length < this.nOutputs) outVec.push(0);
this.dataset.add(inVec, outVec);
}
clearDataset() {
this.dataset.clear();
this.log('Dataset cleared.');
}
randomiseWeights(spread = 0) {
this.storedWeights = this.mlp.getWeights();
this.mlp.drawWeights(spread);
this.weightsRandomised = true;
// Run inference to show effect
const inputWithBias = [...this.inputState, 1.0];
const { output } = this.mlp.getOutput(inputWithBias);
this.outputState = output;
this.log('Weights randomised.');
}
// Add Gaussian noise to weights (for RL exploration)
// spread: 0 = flat noise, 1 = Xavier-scaled per layer
moveWeights(speed, spread = 0) {
this.mlp.moveWeights(speed, spread);
// Run inference to show effect
this.inputUpdated = true;
this.process();
}
train(options = {}) {
// Restore weights if randomised
if (this.weightsRandomised && this.storedWeights) {
this.mlp.setWeights(this.storedWeights);
this.weightsRandomised = false;
}
const features = this.dataset.getFeatures(true); // with bias
const labels = this.dataset.getLabels();
if (features.length === 0 || labels.length === 0) {
this.log('Empty dataset, skipping training.');
return null;
}
this.log('Training...');
this.lastLoss = this.mlp.train(
features,
labels,
this.learningRate,
this.maxIterations,
this.convergenceThreshold,
options
);
const latestHistory = this.mlp.lastTrainingHistory || [];
if (latestHistory.length > 0) {
this.lossHistory.push(...latestHistory);
this.totalTrainingIterations += latestHistory.length;
if (this.lossHistory.length > 1200) {
this.lossHistory = this.lossHistory.slice(this.lossHistory.length - 1200);
}
const runBest = Math.min(...latestHistory);
this.bestLoss = this.bestLoss === null ? runBest : Math.min(this.bestLoss, runBest);
}
// Run inference after training
const inputWithBias = [...this.inputState, 1.0];
const { output } = this.mlp.getOutput(inputWithBias);
this.outputState = output;
this.log(`Training complete. Loss: ${this.lastLoss.toFixed(6)}`);
return this.lastLoss;
}
get exampleCount() {
return this.dataset.size;
}
}