Эксперимент: Moment-Controlled Mutation (#1372)

{ "worked": true, "confidence": 7, "verdict": "Built a NumPy MVP implementing weighted selection, coordinatewise curvature/noise-based diffusion, diversity-floor enforcement, and a fixed-diffusion control. The moment equations were numerically verified: mean and variance-selection relative errors were 0.14% and 0.16%, while the diffusion variance increment matched 2DΔt within 0.00002. Across 8 seeds, moment control improved final mean reward (-0.292 vs -0.341) and stayed closer to target covariance (absolute error 0.0205 vs 0.0448), although fixed diffusion had better mean best-candidate reward (-0.120 vs -0.152); this is a promising small toy signal, not evidence of a general optimizer win.", "metrics": { "baseline": "Fixed D=0.018: mean final reward -0.3411±0.0811, best reward -0.1201±0.0511, covariance-target error 0.0448±0.0439, final covariance 0.1866±0.0439.", "idea": "Moment-controlled diffusion: mean final reward -0.2918±0.0421, best reward -0.1522±0.0371, covariance-target error 0.0205±0.0228, final covariance 0.1523±0.0228, final D 0.0163±0.0049." }, "how_to_run": "/home/maxwelhelp/main/bin/python3 moment_controlled_mutation.py", "files": [ "moment_controlled_mutation.py", "results/summary.json", "results/fixed.jsonl", "results/moment.jsonl", "results/fixed_repetitions.json", "results/moment_repetitions.json" ], "limitations": "This is a noisy synthetic quadratic-vector population, not CIFAR-100, LoRA, or a frozen ViT. The supplied curvature estimator was approximately factor-of-two mis-scaled for an exact Gaussian quadratic: estimate -1.136 versus true curvature -2.3. The MVP follows the supplied estimator formula rather than correcting it. No FLOP-normalized neural-network comparison, hyperparameter sweep, or larger-population test was performed." }