[Meta] ML performance and optimisation

[ma595]

Explore the following issues. Either Daniel/Mandresy or ICCS will do the following:

Optimise training data / ML bias (or better RERUN time)

• Remove LAI and NPP from predictors (from pre-run)
  • If re-run extension less than duration of pre-run -> WIN
• Optimising clustering by selecting pixels that have more PFTs.
• Add more variables that might not be fully equilibrated to the training.
• Vary number of Ncc:
  • Model dependent duration of site-runs ( duration known )
  • Total time of site-runs; Ncc x duration
• How to optimise tradeoff rerun time vs number of pixel for site runs

The issues addressing the above are as follows:

• Add option to remove training on NPP and LAI #71
• Optimise clustering by selecting pixels that have more PFTs #73
• Investigate other variables if NPP and LAI prove to be critical #74
• Routine that scans for optimal value of Ncc #75

[ma595]

Break this down into subtasks and prioritise.

Without LAI and NPP performance was found to be poor. But they are expensive to obtain.

• provide an option to remove these from training.

[dsgoll123]

Here are the sub task:
We need to jointly see which ones we can place in the project time window.

(1) Without LAI and NPP performance was found to be poor. But they are expensive to obtain.

• provide an option to remove these from training. [P1, but easy to do I assume]

(2) Optimising clustering by selecting pixels that have more PFTs. [P0]

• pixel selection is done for each PFT separately, while many pixels contain multiple PFTs
• adjust the clustering to select pixels which can provide information for multipe PFTs in order to reduce the total number of pixels
• the cover fraction of the PFT should be at least 5 (10)% for a given pixel to be considered here
  Untitled document.pdf

(3) If we find LAI and NPP are critical, we could provide additional variables from the same file as predictors ( n = n of response variables ) [P2]

• the associated file contains state variables of ORCHIDEE from early in the spinup (e.g. year 100 vs 3000k when equilbirum is reached)

(4) a routine which allows to scan for optimal Ncc [P0]

• systematically test how Ncc affects stats of ML predictions
• Ncc scale linear with comp demand of ORCHIDEE

[ma595]

Thanks @dsgoll123, I realise that (3) is not high priority but would you be able to provide these additional variables as soon as you can so that we are ready to use them when we have some time? I'll create individual issues so we can put them on the roadmap.

[ma595]

@ma595 asked why Ncc scales linearly with the computational demand of ORCHIDEE. If each pixel can be treated independently, why is this not embarrassingly parallel?

[dsgoll123]

Thanks @dsgoll123, I realise that (3) is not high priority but would you be able to provide these additional variables as soon as you can so that we are ready to use them when we have some time? I'll create individual issues so we can put them on the roadmap.

Yes, we can use the other variables in that file for a test. It is a mixed bag which should contain useful variables but also some without much useful information. Can you work with this or do you want me to pre-select variables?

[dsgoll123]

@ma595 asked why Ncc scales linearly with the computational demand of ORCHIDEE. If each pixel can be treated independently, why is this not embarrassingly parallel?

we are working on it to get there & should be there soon