I tried pulling some data out of Velious/Luclin(no AA) logs to play with validation.
Please note: I have no way to disambiguate movement interrupts. I'm ignoring them for the moment, but presumably that makes this data overestimate interrupts in a potentially non-uniform way. Also, my statistics are rusty so probably don't trust them.
TL;DR - I wouldn't call the proposed function a great fit(potentially just noisy data?), but it does seem notably better than the supplied EQEmu code.
log source
event file
[You must be logged in to view images. Log in or Register.]
summary -
--- Model A: Azxten Formula ---
Expected Successes: 634.42 (63.44%)
Brier Score: 0.2315
Total Error: 8.14%
--- Model A (Azxten) Calibration Table ---
Groups events by their predicted probability.
Pred Prob | Count | Actual % | Expected %
-----------------------------------------------
~ 0.00 | 3 | 0.00% | 2.03%
~ 0.10 | 10 | 30.00% | 13.46%
~ 0.20 | 22 | 22.73% | 22.31%
~ 0.30 | 33 | 36.36% | 31.61%
~ 0.40 | 128 | 28.12% | 39.57%
~ 0.50 | 142 | 43.66% | 48.13%
~ 0.60 | 53 | 69.81% | 58.28%
~ 0.70 | 294 | 59.52% | 69.43%
~ 0.80 | 154 | 64.94% | 78.30%
~ 0.90 | 161 | 76.40% | 88.88%
==================================================
--- Model B: EQEmu Formula ---
Expected Successes: 752.73 (75.27%)
Brier Score: 0.2718
Total Error: 19.97%
--- Model B (EQEmu) Calibration Table ---
Groups events by their predicted probability.
Pred Prob | Count | Actual % | Expected %
-----------------------------------------------
~ 0.50 | 14 | 7.14% | 53.13%
~ 0.60 | 181 | 33.70% | 57.79%
~ 0.70 | 30 | 56.67% | 72.77%
~ 0.80 | 775 | 61.16% | 79.85%