Эксперимент: Discounted Saddle-Gap Controller (#1362)

{ "worked": true, "confidence": 7, "verdict": "Built a projected bilinear min-max MVP with exact saddle-gap evaluation, K-step best-response probing, discounted gap recursion, hysteretic step-size control, and a fixed-step baseline. The recursion matched its direct definition to machine precision and a one-time impulse decayed geometrically with rho, while the controller reduced mean post-change gap from 10.36 to 6.21 across three seeds. The effect is real in this toy nonstationary game, but the controller achieved it by repeatedly shrinking the step size to about 1.68e-6, so this is only a promising stability signal rather than evidence of broad GAN utility.", "metrics": { "baseline": "Mean post-change exact saddle gap 10.362; last-300 gap 10.400; fixed step 0.055; 3 seeds.", "idea": "Mean post-change exact saddle gap 6.206; last-300 gap 6.207; final step 1.68e-6; 15 downward events; 3 seeds. Recursion and impulse-decay max errors were 0 and 1.39e-17." }, "how_to_run": "/home/maxwelhelp/main/bin/python3 experiment.py", "files": [ "experiment.py", "README.txt" ], "limitations": "Only a scalar projected bilinear game was tested. No neural GAN, real data, momentum-state clearing, equal-FLOP comparison, baseline loss-EMA controller, extragradient, or sample-quality metric was evaluated; probe-gap quality and the severe learning-rate collapse require further study." }