js/neuralNetwork.js

/**
 * Lightweight feedforward neural network for entity decision-making.
 * Supports arbitrary topology, weight serialization, mutation, and
 * activation visualization for the HUD neural activity overlay.
 */
export class NeuralNetwork {
  /**
   * @param {number[]} topology - Layer sizes, e.g. [12, 10, 6]
   */
  constructor(topology) {
    this.topology = topology;
    this.weights = [];
    this.biases = [];
    this.activations = [];
    this._initWeights();
  }

  /** Xavier-style initialization for stable gradient flow */
  _initWeights() {
    for (let i = 1; i < this.topology.length; i++) {
      const fanIn = this.topology[i - 1];
      const fanOut = this.topology[i];
      const scale = Math.sqrt(2 / (fanIn + fanOut));
      const layerW = new Array(fanOut);
      const layerB = new Array(fanOut);
      for (let j = 0; j < fanOut; j++) {
        layerW[j] = new Array(fanIn);
        for (let k = 0; k < fanIn; k++) {
          layerW[j][k] = (Math.random() * 2 - 1) * scale;
        }
        layerB[j] = (Math.random() * 2 - 1) * 0.1;
      }
      this.weights.push(layerW);
      this.biases.push(layerB);
    }
  }

  /**
   * Run a forward pass through the network.
   * @param {number[]} inputs - Input activations
   * @returns {number[]} Output activations
   */
  forward(inputs) {
    let current = inputs;
    this.activations = [inputs.slice()];

    for (let layer = 0; layer < this.weights.length; layer++) {
      const w = this.weights[layer];
      const b = this.biases[layer];
      const numNeurons = w.length;
      const next = new Array(numNeurons);
      const isOutput = layer === this.weights.length - 1;

      for (let j = 0; j < numNeurons; j++) {
        let sum = b[j];
        const wj = w[j];
        for (let k = 0; k < current.length; k++) {
          sum += wj[k] * current[k];
        }
        next[j] = Math.tanh(sum);
      }

      current = next;
      this.activations.push(current.slice());
    }

    return current;
  }

  /** Count total trainable parameters */
  get totalWeights() {
    let count = 0;
    for (let i = 0; i < this.weights.length; i++) {
      count += this.weights[i].length * this.weights[i][0].length;
      count += this.biases[i].length;
    }
    return count;
  }

  /** Flatten all weights + biases into a single array */
  serialize() {
    const flat = new Array(this.totalWeights);
    let idx = 0;
    for (let i = 0; i < this.weights.length; i++) {
      const w = this.weights[i];
      for (let j = 0; j < w.length; j++) {
        for (let k = 0; k < w[j].length; k++) {
          flat[idx++] = w[j][k];
        }
      }
      const b = this.biases[i];
      for (let j = 0; j < b.length; j++) {
        flat[idx++] = b[j];
      }
    }
    return flat;
  }

  /** Restore weights from a flat array */
  deserialize(flat) {
    let idx = 0;
    for (let i = 0; i < this.weights.length; i++) {
      const w = this.weights[i];
      for (let j = 0; j < w.length; j++) {
        for (let k = 0; k < w[j].length; k++) {
          w[j][k] = flat[idx++];
        }
      }
      const b = this.biases[i];
      for (let j = 0; j < b.length; j++) {
        b[j] = flat[idx++];
      }
    }
  }

  /** Create an exact copy of this network */
  clone() {
    const nn = new NeuralNetwork(this.topology.slice());
    nn.deserialize(this.serialize());
    return nn;
  }

  /**
   * Apply Gaussian mutations to a fraction of weights.
   * @param {number} rate - Probability of mutating each weight (0-1)
   * @param {number} magnitude - Standard deviation of mutation noise
   */
  mutate(rate, magnitude) {
    for (let i = 0; i < this.weights.length; i++) {
      const w = this.weights[i];
      for (let j = 0; j < w.length; j++) {
        for (let k = 0; k < w[j].length; k++) {
          if (Math.random() < rate) {
            w[j][k] += gaussianNoise() * magnitude;
          }
        }
      }
      const b = this.biases[i];
      for (let j = 0; j < b.length; j++) {
        if (Math.random() < rate) {
          b[j] += gaussianNoise() * magnitude;
        }
      }
    }
  }

  /**
   * Uniform crossover between two networks of the same topology.
   * @param {NeuralNetwork} other
   * @returns {NeuralNetwork} child network
   */
  crossover(other) {
    const child = new NeuralNetwork(this.topology.slice());
    const w1 = this.serialize();
    const w2 = other.serialize();
    const childW = new Array(w1.length);

    const crossPoint = Math.floor(Math.random() * w1.length);
    for (let i = 0; i < w1.length; i++) {
      childW[i] = i < crossPoint ? w1[i] : w2[i];
    }
    child.deserialize(childW);
    return child;
  }

  /**
   * Get the maximum absolute activation across all layers.
   * Useful for visualization intensity scaling.
   */
  getMaxActivation() {
    let max = 0;
    for (const layer of this.activations) {
      for (const v of layer) {
        const abs = Math.abs(v);
        if (abs > max) max = abs;
      }
    }
    return max;
  }

  /**
   * Render a mini neural network diagram to a canvas context.
   * @param {CanvasRenderingContext2D} ctx
   * @param {number} x - Top-left x
   * @param {number} y - Top-left y
   * @param {number} w - Width
   * @param {number} h - Height
   */
  renderMini(ctx, x, y, w, h) {
    const layers = this.topology.length;
    const layerSpacing = w / (layers - 1);

    ctx.lineWidth = 0.5;
    for (let l = 0; l < this.weights.length; l++) {
      const fromCount = this.topology[l];
      const toCount = this.topology[l + 1];
      const fromX = x + l * layerSpacing;
      const toX = x + (l + 1) * layerSpacing;

      for (let j = 0; j < toCount; j++) {
        const toY = y + (j + 0.5) * (h / toCount);
        for (let k = 0; k < fromCount; k++) {
          const fromY = y + (k + 0.5) * (h / fromCount);
          const weight = this.weights[l][j][k];
          const alpha = Math.min(Math.abs(weight) * 0.5, 0.6);
          ctx.strokeStyle =
            weight > 0
              ? `rgba(0, 240, 255, ${alpha})`
              : `rgba(255, 51, 102, ${alpha})`;
          ctx.beginPath();
          ctx.moveTo(fromX, fromY);
          ctx.lineTo(toX, toY);
          ctx.stroke();
        }
      }
    }

    for (let l = 0; l < layers; l++) {
      const count = this.topology[l];
      const lx = x + l * layerSpacing;
      const act = this.activations[l] || [];

      for (let n = 0; n < count; n++) {
        const ny = y + (n + 0.5) * (h / count);
        const val = act[n] || 0;
        const intensity = Math.abs(val);
        const r = Math.max(2, 4 - layers * 0.3);

        ctx.beginPath();
        ctx.arc(lx, ny, r, 0, Math.PI * 2);
        ctx.fillStyle =
          val > 0
            ? `rgba(0, 240, 255, ${0.3 + intensity * 0.7})`
            : `rgba(255, 51, 102, ${0.3 + intensity * 0.7})`;
        ctx.fill();
      }
    }
  }
}

/** Gaussian noise using Box-Muller transform */
function gaussianNoise() {
  let u = 0,
    v = 0;
  while (u === 0) u = Math.random();
  while (v === 0) v = Math.random();
  return Math.sqrt(-2.0 * Math.log(u)) * Math.cos(2.0 * Math.PI * v);
}

/**
 * QuantumDecisionLayer - probabilistic branching for elite entities.
 * Samples from a probability distribution over actions rather than
 * taking the deterministic argmax.
 */
export class QuantumDecisionLayer {
  /**
   * @param {number} temperature - Controls exploration vs exploitation.
   *   Higher = more random, lower = more greedy.
   */
  constructor(temperature = 1.0) {
    this.temperature = temperature;
  }

  /**
   * Apply softmax with temperature to raw outputs and sample.
   * @param {number[]} logits - Raw network outputs
   * @returns {{ action: number, probabilities: number[] }}
   */
  sample(logits) {
    const t = this.temperature;
    const maxLogit = Math.max(...logits);
    const exps = logits.map((l) => Math.exp((l - maxLogit) / t));
    const sum = exps.reduce((a, b) => a + b, 0);
    const probs = exps.map((e) => e / sum);

    const r = Math.random();
    let cumulative = 0;
    for (let i = 0; i < probs.length; i++) {
      cumulative += probs[i];
      if (r <= cumulative) {
        return { action: i, probabilities: probs };
      }
    }
    return { action: probs.length - 1, probabilities: probs };
  }
}