Эксперимент: Bounded Telegraph Exploration for Optimizers (#1106)

{ "worked": false, "confidence": 9, "verdict": "Built and numerically tested bounded telegraph exploration with exact discrete relaxation. The strict bound held in every run (maximum observed |u|/u_max=0.999999), and the variance law matched reasonably well overall. However, the supplied covariance formula is invalid because it predicts C(0)=0; the corrected covariance matched simulation much better, while kurtosis scaling was noisy and the quadratic optimizer showed no meaningful win over matched-variance OU noise.", "metrics": { "baseline": "Matched-variance Gaussian OU: final quadratic loss 0.02503; average last-500 loss 0.03497.", "idea": "Telegraph: final quadratic loss 0.02832; average last-500 loss 0.03463; maximum observed bound ratio 0.999999; covariance RMSE 0.0355 for the corrected formula versus 0.2371 for the supplied formula." }, "how_to_run": "/home/maxwelhelp/main/bin/python3 telegraph_experiment.py", "files": [ "telegraph_experiment.py", "results.json", "run_output.txt" ], "limitations": "Only a toy quadratic optimizer was tested, not CIFAR-10, an MLP, or ResNet. No wall-clock/FLOP or generalization measurement was performed, and the finite Monte Carlo horizon made the kurtosis sweep noisy." }