prototype: GA evolution spike -- sin(x) prediction validates pipeline (#66)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-08-24 08:14:32 +02:00
parent b7b10af811
commit eee48dec38
1 changed files with 111 additions and 0 deletions
@@ -0,0 +1,111 @@
+## ga_spike.nim — throwaway GA neuroevolution spike: evolve ANN to predict sin(x)
+## Validates: population init, fitness eval, truncation selection, gaussian
+## mutation on all weights, single-elite preservation, champion tracking.
+## Issue #66.
+
+import std/[math, random, algorithm, strformat]
+
+const
+  InputDim  = 10   # sliding window of 10 past sin values
+  HiddenDim = 4
+  OutputDim = 1
+  NumWeights = (InputDim * HiddenDim + HiddenDim) +  # W1 + b1
+               (HiddenDim * OutputDim + OutputDim)    # W2 + b2  = 49
+  PopSize    = 200
+  Generations = 200
+  TopFrac    = 0.2  # keep top 20%
+  Sigma      = 0.01 # gaussian mutation sigma on ALL weights
+  EvalPoints = 50   # fitness eval sample size
+
+type Individual = object
+  weights: seq[float64]
+  fitness: float64
+
+proc forward(w: seq[float64]; input: array[InputDim, float64]): float64 =
+  ## 10->4->1 tanh ANN, flat weight layout: W1[40], b1[4], W2[4], b2[1]
+  var hidden: array[HiddenDim, float64]
+  for h in 0..<HiddenDim:
+    var s = w[InputDim * HiddenDim + h]  # b1[h]
+    for i in 0..<InputDim:
+      s += w[h * InputDim + i] * input[i]  # W1[h,i]
+    hidden[h] = tanh(s)
+  var s = w[InputDim * HiddenDim + HiddenDim + HiddenDim]  # b2[0]
+  for h in 0..<HiddenDim:
+    s += w[InputDim * HiddenDim + HiddenDim + h] * hidden[h]  # W2[h]
+  result = tanh(s)
+
+proc evaluate(ind: var Individual; rng: var Rand) =
+  ## Fitness = -MSE on EvalPoints random samples of sin prediction.
+  var mse = 0.0
+  for _ in 0..<EvalPoints:
+    let x = rng.rand(0.0 .. 20.0 * PI)
+    var input: array[InputDim, float64]
+    for i in 0..<InputDim:
+      input[i] = sin(x - float64(InputDim - 1 - i))
+    let target = sin(x)
+    let pred = forward(ind.weights, input)
+    mse += (pred - target) * (pred - target)
+  ind.fitness = -mse / EvalPoints.float64
+
+proc mutate(ind: var Individual; rng: var Rand) =
+  ## Additive gaussian on ALL weights, per GA-parameters research.
+  for i in 0..<ind.weights.len:
+    ind.weights[i] += rng.gauss(0.0, Sigma)
+
+when isMainModule:
+  var rng = initRand(42)
+
+  # Init population: random weights in [-0.5, 0.5]
+  var pop = newSeq[Individual](PopSize)
+  for i in 0..<PopSize:
+    pop[i].weights = newSeq[float64](NumWeights)
+    for j in 0..<NumWeights:
+      pop[i].weights[j] = rng.rand(-0.5 .. 0.5)
+
+  echo &"GA spike: {PopSize} individuals, {NumWeights} weights, {Generations} gens, sigma={Sigma}"
+
+  for gen in 0..<Generations:
+    # Evaluate
+    for i in 0..<PopSize:
+      pop[i].evaluate(rng)
+
+    # Sort descending by fitness (higher = better)
+    pop.sort(proc(a, b: Individual): int = cmp(b.fitness, a.fitness))
+
+    if gen mod 20 == 0 or gen == Generations - 1:
+      echo &"  gen {gen:3d}  best_fitness={pop[0].fitness:.6f}  (MSE={-pop[0].fitness:.6f})"
+
+    # Selection: keep top 20%
+    let nParents = max(1, int(PopSize.float64 * TopFrac))
+
+    # Next generation: elite(1) + mutated children from parents
+    var next = newSeq[Individual](PopSize)
+    next[0] = pop[0]  # single elite, unchanged
+    for i in 1..<PopSize:
+      let parent = rng.rand(0..<nParents)
+      next[i].weights = pop[parent].weights  # clone parent
+      next[i].mutate(rng)
+    pop = next
+
+  # Final evaluation of champion
+  pop[0].evaluate(rng)
+  echo &"\nChampion fitness: {pop[0].fitness:.6f}  (MSE={-pop[0].fitness:.6f})"
+
+  # Print predictions vs actual
+  echo "\nPredictions (sample):"
+  var totalErr = 0.0
+  let nSamples = 20
+  for i in 0..<nSamples:
+    let x = float64(i) * 1.0
+    var input: array[InputDim, float64]
+    for j in 0..<InputDim:
+      input[j] = sin(x - float64(InputDim - 1 - j))
+    let target = sin(x)
+    let pred = forward(pop[0].weights, input)
+    totalErr += abs(pred - target)
+    echo &"  x={x:5.1f}  sin(x)={target:+.4f}  pred={pred:+.4f}  err={abs(pred-target):.4f}"
+
+  let avgErr = totalErr / nSamples.float64
+  echo &"\nAverage absolute error: {avgErr:.4f}"
+  assert avgErr < 0.1, &"Champion avg error {avgErr:.4f} >= 0.1 — evolution did not converge"
+  echo "PASS: assert avgErr < 0.1"