SimonGoring
commented
11 years ago Okay. This sounds good. I'm running randomForest on my work computer right now, while I'm out of town just so see how it goes. I'll do a push up to github once I get a few passes finished. @jssalonen can you take the code for boot_rfor and change it to a rough brt code? Just so that we've got a semi-complete set of methods written in? Even if it doesn't have the parameterization, at least we'll have something there.
It may not be necessary to develop separate sets of BRT parameters for different sample sizes, as we're in a vast majority of cases looking at a loss of less than 500 sites, and mostly just a couple hundred. Instead, let's look for a good set of parameters that performs well with all commonly expected calibration sets. We should be able to find a parameter combo that performs consistently enough for the purposes of this paper (observing the general behaviour of BRTs in no-analogue conditions).
We'll have to do our best to test the parameter combo VERY well - we certainly don't want to have to repeat the run! I think first we'll check a bootstrapped sample from the whole data-set. Then we might take a small selection of samples from different sub-regions, drop all samples within sq. chord dist. 1.0, and test with bootstrapped samples of those sets. This'll give us some idea about what works with least-analogous cal sets.
Sakari & Miska will start looking into this.