Эксперимент: Minimum-motion curvature-targeted preconditioner (#1365)

{ "worked": false, "confidence": 9, "verdict": "Built a diagonal log-metric receding-horizon controller with terminal generalized-condition penalty and affine-invariant kinetic surrogate. The mathematical checks passed to numerical precision, and the controller reduced metric jumps by about 36% versus the Adam-style EMA control. However, across five paired noisy-quadratic runs it had worse mean conditioning, generally worse final loss, slightly higher gradient variance, and roughly 8x runtime, exceeding the claimed overhead budget; therefore the optimization win was not observed.", "metrics": { "baseline": "Across seeds: final loss 0.0158-0.0320, mean generalized condition 1.38-1.67, metric jump 0.195-0.205, gradient variance 2.75-3.12, runtime 0.084-0.122 s.", "idea": "Across seeds: final loss 0.0152-0.0417, mean generalized condition 4.64-5.51, metric jump 0.125-0.147, gradient variance 2.81-3.21, runtime 0.691-0.749 s. Math errors were approximately 2.2e-16 for diagonal affine distance, 2.2e-16 for affine invariance, and 6.9e-18 for the constant-speed kinetic identity." }, "how_to_run": "/home/maxwelhelp/main/bin/python3 experiment.py", "files": [ "experiment.py", "results.json" ], "limitations": "Only a small diagonal noisy quadratic was tested; no MLP, MNIST/CIFAR data, block-SPD metrics, Hessian-vector products, GPU execution, or wall-clock comparison on a neural network was evaluated. The curvature estimate was synthetic noisy diagonal curvature rather than an empirical neural-network Hessian or Fisher estimate." }