Heavy-Tail Path-Adaptive Optimizer Pool / README.md
Mechanism failed
Heavy-Tail Path-Adaptive Optimizer Pool MVP
optimizer_pool.py implements restarted diagonal AdaGrad experts with geometric horizons and the centered variance-normalized exponential meta-update from the idea.
Run:
/home/maxwelhelp/main/bin/python3 optimizer_pool.py
The script verifies two analytic predictions by sweeps:
- For a constant scalar gradient, cumulative AdaGrad displacement scales as
sqrt(H); the asymptotic prediction is2 alpha sqrt(H). - For two fixed-loss experts with gap
delta, the centered meta log-odds scale as4 beta sqrt(T)under the stated per-expert variance accumulators.
It also reports a deliberately falsifiable dynamic-recovery sweep and a shifted quadratic comparison. The latter is exploratory only; the pool uses multiple expert updates and is not a claim of equal-FLOP superiority.