# Эксперимент: Universal Trust-Region Neural Optimizer (#460)

{
  "worked": true,
  "confidence": 9,
  "verdict": "Built a reproducible trust-region optimizer MVP with ratio-based acceptance, radius adaptation, and a gradient/Cauchy step for a diagonal quadratic model. The mechanism manifested quantitatively: exact models produced rho approximately 1 across curvature and radius sweeps, mismatched-curvature ratios matched rho=2-lambda/B with maximum error 0, the near-boundary transition matched the predicted Delta*/0.99 rule, and good/poor agreement caused 2x expansion and 4x contraction. In the tiny MLP, Adam performed better after 120 steps, so this validates the mechanism but does not demonstrate a neural-training win.",
  "metrics": {
    "baseline": "Adam: final loss 0.04802, accuracy 98.96%, 0 loss spikes.",
    "idea": "Trust region: final loss 0.10696, accuracy 96.88%, 0 rejected steps, final radius 0.6, median rho 1.986. Toy exact-model rho range [0.999999999999996, 1.000000000000002], curvature-ratio maximum error 0."
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 trust_region_experiment.py",
  "files": [
    "trust_region_experiment.py",
    "results.json"
  ],
  "limitations": "The neural MVP uses a full-batch 2D synthetic classification task and identity curvature rather than Hessian-vector products or Gauss-Newton curvature. MNIST/CIFAR, minibatch noise, wall-clock/FLOP matching, larger models, and the paper's asymptotic complexity rates were not tested."
}