# Эксперимент: Noise-Whitened Trajectory-KL Policy Regularization (#506)

{
  "worked": true,
  "confidence": 8,
  "verdict": "Built a linear-Gaussian MVP implementing the inverse-covariance drift penalty, analytic optimizer, finite-difference math check, parameter sweeps, and a fixed-seed PyTorch policy comparison. The mechanism manifested: trajectory KL decreased monotonically with lambda, large-lambda scaling approached the predicted noise-dependent limit with 14.7% maximum relative error, and measured curvature exactly matched the predicted ratio of 4.0. The idea reduced trajectory KL by 77.1% versus action regularization, while task loss increased slightly, demonstrating the intended preservation/task tradeoff rather than an unconditional task win.",
  "metrics": {
    "baseline": "Action-norm regularization: task half-MSE 0.39499, trajectory KL 0.16137, action penalty 0.04937.",
    "idea": "Noise-whitened drift regularization: task half-MSE 0.41197, trajectory KL 0.03702, action penalty 0.11427; toy KL decreased from 8.5 at lambda=0 to 0.000789 at lambda=100; curvature ratio observed/predicted 4.0/4.0; asymptotic lambda*u relative error 14.7%."
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 experiment.py",
  "files": [
    "experiment.py",
    "results.json"
  ],
  "limitations": "Only a synthetic linear-Gaussian setting was tested; no Pendulum/PPO, learned drift or diffusion model, finite-horizon rollout KL estimation, stochastic policy distributions, model misspecification, or broad lambda/task-performance sweep was evaluated."
}