brianlangseth-NOAA
commented
1 year ago Decisions made to improve on the model so as to have more 100 datasets more reasonable include -increasing bounds, especially for sel_beta3 and ...4. This change became moot when we changed problematic fleet selectivities from dbl logistic to flat topped logistic. -Change fleets 4 and 6 from domed to flat-topped logistic. Model was estimating a flat top in these cases and causing parameter bound issues. We know they are not flat topped, but spatial or temporal dynamics that we're missing are clearly pushing the model to flat-topped. -Increase sample size for fleet 7 to 25, just like fleet 5. This change fixed a wonky looking selectivity pattern that was being estimated for fleet 7, as it did previously for fleet 5.
These improved performance. Previously ~50 of the 100 datasets were hitting bounds or had high gradients, whereas after these chances, ~12 of the 100 datasets were hitting bounds or had high gradients.
Changes made in commit e9db539, which closes this issue
In the process of trying to get 100 reasonable, single area, fits completed, I discovered that our initial, single dataset fit had several parameters hitting bounds. These were mostly related to sel_beta3 and sel_beta4, but there could be others. Generally, we need to setup a parameter bound check for all parameters. For the simulation exercise, we should be as vigilante as possible, with special attention to selectivity and q parameters. The data_loading scripts can be easily amended to change the bounds.