Эксперимент: Global-Local Koopman Latent Dynamics (#1335)

{ "worked": false, "confidence": 9, "verdict": "Built a fixed-seed global/local Koopman latent predictor with nonlinear encoders, separate learned global/local linear operators, and a joint decoder, plus a GRU baseline. The block update reproduced exactly (maximum numerical discrepancy 0), and both operator blocks had spectral radius below 1 with zero-input norm ratio 0.534, confirming the claimed stable recursive behavior. However, prediction accuracy did not improve: at 20 steps the idea reached MSE 0.05286 versus 0.02278 for the GRU baseline, and it was also worse at one and ten steps; therefore no overall win was observed.", "metrics": { "baseline": "20-step MSE 0.0227833; 10-step MSE 0.0169388; 1-step MSE 0.0129002; 2,760 parameters; training 19.50 s", "idea": "20-step MSE 0.0528566; 10-step MSE 0.0799640; 1-step MSE 0.167647; 2,066 parameters; training 24.79 s; learned global/local spectral radii 0.98087/0.98813" }, "how_to_run": "/home/maxwelhelp/main/bin/python3 run_experiment.py", "files": [ "run_experiment.py", "results.json" ], "limitations": "Only a synthetic 12-dimensional sequence system and one fixed train/test split were tested; no Moving-MNIST, real robot data, controls, altered disturbances, FLOP-normalized benchmark, wall-clock rollout benchmark, or hyperparameter sweep was performed. The experiment uses only the first observation as encoder context rather than a multi-frame history." }