Эксперимент: CVaR-tail active residual correction (#1105)

{ "worked": true, "confidence": 8, "verdict": "Built a self-contained NumPy MVP for CVaR-tail active residual correction, including empirical VaR/CVaR, ensemble tail-region selection, bootstrap residual uncertainty, ridge correction, and three acquisition policies. All three mechanism checks matched quantitative predictions: Rockafellar optimization recovered the 0.99 quantile within 2.52e-4, ridge shrinkage matched 1/(1+lambda) exactly, and tail-region coverage matched 2Phi(epsilon)-1 within 1.45e-3. The tail-residual policy improved over global uncertainty for CVaR error at every measured budget after initialization, but did not consistently beat random sampling, so a broad sample-efficiency win was not demonstrated.", "metrics": { "baseline": "Random CVaR absolute error by calls: 0.00583, 0.000732, 0.0000626, 0.000311, 0.00170 at 24, 40, 56, 72, 88 calls; global uncertainty: 0.0286, 0.00492, 0.00398, 0.00335, 0.00299.", "idea": "Tail-residual CVaR absolute error: 0.00879, 0.00191, 0.000964, 0.00116, 0.00118 at 24, 40, 56, 72, 88 calls; true CVaR=1.65539. Mechanism errors: CVaR eta 2.52e-4, ridge ratio 0, tail coverage 1.45e-3 maximum." }, "how_to_run": "/home/maxwelhelp/main/bin/python3 experiment.py", "files": [ "experiment.py", "results.json", "run.log" ], "limitations": "The surrogate ensemble is analytic rather than neural, the synthetic problem is one-dimensional, and the policy comparison used one trajectory per policy rather than repeated independent trials. GPU execution, high-dimensional inputs, actual MLP training, downstream constrained optimization, and statistical confidence intervals were not tested." }