re-scaling at each iteration does wonders

[lorenzozanisi]

I am now unscaling all datasets at the end of each iteration, adding the (unscaled) candidates to the (unscaled) train, then create a new scaler based on this new training set and rescale everything else. The improvement is crazy. As usual, train on the left, test on the right.

@zcapjdb , following from #78 I was getting these low losses because I was doing this new re-scaling method, but I forgot...

[zcapjdb]

Great! I assume you made those changes in your branch which isn't really compatible with main, could you copy those changes into a branch that can be merged into main

[lorenzozanisi]

Two things to note:

• even with the standard method, we can start from only 2,000 points, which is nothing, and scale up to 20,000 and reach a decent mean MSE. But there's a huge scatter in different models (some get to 0.1, some to 0.17 which is not great).
• Instead, with the new rescaler, we improve on the mean by a whopping factor of FIVE (!)
• With the rescaler we have a much smaller scatter between models (although in percentage it might be similar, but it's the absolute the matters for us)
• the pipeline converges so much faster (only about 5 iterations)

@zcapjdb I'm running this again with 20,000 inputs for 20 iterations, just to check whether this behaviour is stable, and then will merge to main