search

GLEON / MilleLacsMystery

Bass/Walleye modeling of Mille Lacs, MN
0 stars 3 forks source link

Run GLM and remove heat budget bias for the lake #4

Open jordansread opened 7 years ago

lawinslow commented 7 years ago

Ok, running with preliminary values and non-time-varying Kd just to get a handle on the optimization. I think I've figured out the magic to getting GLM to properly optimize. Has a lot to do with telling optim how much it should be modifying parameters so it doesn't make huge, awkward leaps around the likelihood surface. Currently using RMSE as the fit metric.

Default parameters have RMSE ~1.7 degC. We'll see where it goes from here.

jordansread commented 7 years ago

Not a bad starting point (1.7°)

You are using all of the temp data? Including all of the new data Gretchen dug up? Assuming it is a lot more than we had for validation in NECSC.

gjahansen2 commented 7 years ago

It is a lot, lot, lot more than we had before.

On Wed, Apr 5, 2017 at 10:31 AM, Jordan S Read notifications@github.com wrote:

Not a bad starting point (1.7°)

You are using all of the temp data? Including all of the new data Gretchen dug up? Assuming it is a lot more than we had for validation in NECSC.

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/GLEON/MilleLacsMystery/issues/4#issuecomment-291899665, or mute the thread https://github.com/notifications/unsubscribe-auth/ASiU6sH9zWmizW2yE5nm26Co8pbyjSTsks5rs7PpgaJpZM4Mojyk .

-- Gretchen Hansen https://gretchenhansen.squarespace.com

lawinslow commented 7 years ago

Finished. Final RMSE: 1.26 deg C image

Hmmm, I may drop kw_factor out of there. That seems to have gone weird places (should be 1 +/- a little).

Re-running.

gjahansen2 commented 7 years ago

what is kw_factor? what is counts? Does convergence=0 mean something bad?

lawinslow commented 7 years ago

0 is good for convergence. Anything not zero is usually bad. 0 tends to mean it found a local minimum, which is all we can guarantee with the method I'm using.

kw_factor was a Kd multiplier i had in there. I was using for other lakes. Tweaks our estimate of light attenuation. Something near 1 is fine, means it made a small adjustment to Kd. In this case, we know it went somewhere weird because it was saying Kd is ~10x smaller than the value I supplied, which we know isn't true.

lawinslow commented 7 years ago

Weird. It really doesn't like the clarity value I'm using right now. Maybe the variable Kd will have a really large impact then given that with a fixed Kd, I can't really improve RMSE very much.

image

For RMSE, it definitely seems to think the lake should be clearer than the value I"m giving it.

gjahansen2 commented 7 years ago

what value are you giving it?

lawinslow commented 7 years ago

Kd of 0.5, which is probably high, but not high enough to believe a kd_factor (multiplier) of 0.09 as the first run gave me. I'll work on getting the variable Kd run going and see how that optimization turns out.

gjahansen2 commented 7 years ago

It would be nice to see errors over time, too., under the constant kd scenario.