Assign random number seed independent of num_threads

[sebp]

Addresses my comment in #1262 (comment)

ForestTrainer::train_trees calls split_sequence to split the number of trees to grow into batches depending on the specified number of threads. However, when assigning a seed to grow an individual tree, only the batch index is considered, but not the tree index. Therefore, different values for the number of threads will yield different results, despite the user having specified the same (global) random number seed.

The patch initializes the random number generator based on the user provided seed and the index of the tree (start + i). Thus, each tree will be assigned the same seed, independent of the specified number of threads.

[erikcs]

Thanks @sebp, you are right this would be the preferred approach.

But, since this is how grf has worked since inception and changing it would be a breaking change, we decided to keep it that way:

Therefore, in order to ensure consistent results, we provide the following recommendations.

• Make sure arguments seed and num.threads are the same across platforms

[sebp]

I understand that you want to preserve backwards compatibility. If you someday release a new major version, it would be nice include this fix.

[erikcs]

Thanks, will do.