Эксперимент: Polynomial-Lyapunov Training Controller (#1318)

{ "worked": false, "confidence": 9, "verdict": "Built a polynomial-Lyapunov controller with monitoring-batch rejection, EMA-smoothed c-hat, adaptive learning rate, and an exact comparison-law sanity check. The toy ODE discretization matched c=0.25 closely and produced algebraic log-log slopes near the predicted -1/m for m=1 and m=2. However, in the digits MLP experiment the controller was worse than fixed-step SGD for two of three seeds; m=2 improved once but not reproducibly, while rejection rates were high and learning rates often collapsed.", "metrics": { "baseline": "Seed 3126 test loss 0.1824; seed 7 0.1987; seed 42 0.2072; fixed eta=0.12, no rejections.", "idea": "Best observed m=2 test loss 0.1622 on seed 3126, but 0.3208 on seed 7 and 0.2655 on seed 42; rejection rates 45.6%-50.6% and final eta frequently 0.0001." }, "how_to_run": "SEED=3126 /home/maxwelhelp/main/bin/python3 experiment.py", "files": [ "experiment.py", "results.json" ], "limitations": "Only a small sklearn digits MLP was tested, not MNIST or CIFAR-10/ResNet; only three seeds and 500 steps were used. c_star was manually calibrated separately for each m, no wall-clock/FLOP comparison or learning-rate boundary sweep was performed, and the asymptotic slope claim was verified on the exact toy ODE rather than reliably measured in neural-network training." }