Spaces:

---

calibrationcomp

/

calibration_benchmark

like 0

Sleeping

Contains:

• Adam results data with parameters: (learning rate, mom & grad^2 moving average weights, and regularization guard, number of iterations before determining failure)
  $$\alpha =0.001,{\beta }_m=0.9,{\beta }_v=0.999,\epsilon ={10}^{-8},N_{iter}=10,000\backslash alpha\; =\; 0.001,\; \backslash beta\_m\; =\; 0.9,\; \backslash beta\_v\; =\; 0.999,\; \backslash varepsilon\; =\; 10^\{-8\},\; N\_\{iter\}=10,000$$

• NB: Tried also with larger learning rate (0.01) with no performance difference. (could play endlessly, but will keep fixed as a first go).

• Added this to optimization leaderboard.

• Extended functionality to handle that the vector adam run failed for every restart

TODO:

• Add Levenberg Marquardt results too

Also now contains:

• Levenberg Marquardt algorithm for l63, l96_const, l96_vec (does a better than adam) but still essentially fails on all experiments (failure rate for targets >60%)
• uses adaptive lambda and pivoted QR solve for maximum robustness (as implemented in MINPACK)

Leaderboard improvements:

• the suitability table is grouped by algorithm type (better "block" readability)
• the display colors are now consistent across all leaderboards and experiments, and vline for single ensemble runs (e.g. easier to see UKI)
• the failure rate table is added to the optimization experiment