Эксперимент: Chernoff-Tied Neural Evolution (#1232)

{ "worked": false, "confidence": 8, "verdict": "Built a shared-parameter Euler residual neural evolution block and an untied depth-20 baseline for a periodic advection-diffusion system. The generator check was accurate to about 1e-6 numerical precision, and measured perturbation errors stayed below the discrete Gronwall bounds. However, at t=1 the tied model had RMSE 0.0295 versus 0.0211 for the untied baseline, so this MVP does not demonstrate an accuracy win; tied extrapolated rollouts remained stable but were not compared against a valid untied model beyond its trained depth.", "metrics": { "baseline": "Untied depth-20: 11,040 parameters; t=1 RMSE 0.0211053", "idea": "Tied 20-step block: 552 parameters; t=1 RMSE 0.0295355, t=2 0.0223241, t=5 0.0253307, t=10 0.0334833; generator errors 1.14e-7 to 9.90e-7 and perturbation final errors below bounds" }, "how_to_run": "python3 experiment.py", "files": [ "experiment.py", "results.json" ], "limitations": "Only a tiny 8-dimensional linear advection-diffusion toy system was tested. Training used a short 500-iteration budget, no spectral normalization or learned step size, and the untied baseline was trained only for t=1, so long-horizon baseline comparison and FLOP benchmarking were not performed." }