Hi Mr Popov,
Just wondering if there's any way to generate more "consistency" across strategies in Exp.Gen...difficult to explain, but hopefully the case in point below will help.

I assume it's starting with some kind of random number to begin generating strategies, and so my case in point:
I launched 10 Exp.Gen's simultaneously (scripted so admittedly a minor delay between each launch) on the same 1H CADJPY dataset using the same parameters, running the generation for 45 minutes.

I then uploaded the dataset, and the individual collections into EAS, exported a portfolio for each collection and forward tested it in MT5 for 96 bars, starting 1 bar after the final bar from the dataset (so, dataset was 09:00 GMT, Portfolio EA launched at 10:00 GMT).

The results were actually quite dramatically different:

    Profit    Win Rate    Avg Profit    Avg Loss
0    321.69    54.84%    13.28    4.64
1    206.68    52.17%    12.71    4.47
2    138.81    43.08%    11.63    5.05
3    267.21    43.08%    14.34    3.63
4    278.54    49.30%    12.59    4.5
5    200.24    50.72%    10.01    4.42
6    268.91    52.83%    13.62    4.5
7    300.49    62.79%    13.86    4.62
8    171.95    40.30%    12.27    3.98
9    160.01    53.85%      7.27    4.03
*    472.85    47.52%    14.81    4.49 (merged using Exp.Gen)
*    783.38    52.88%    18.84    5.16 (merged using EAS Validator)
*    337.47    46.07%    13.62    4.6 (single generation)

TP was fixed at 150, SL at 25, 360 bars of data, and using Reverse direction, lot size of 0.02.  Also using M/Carlo (same settings as generation)...happy to share my settings file.  0-9 were all generated simultaneously.

There's quite a big difference across all metrics for the various portfolios.  They all started at basically the same time, and each had about 20million calculations.

I re-ran for a couple of other dates and the same thing occurred.

So I just wondered if there's a way we can create "more" consistency across the generation process...these seem pretty random and therefore results would be more luck than anything else?  I understand the randomness of things, and the volume of permutations and combinations of things, etc.

Couple of my own thoughts/questions...is it better to have one process running to generate strategies - on my computer it seems to be able to go through about 55million calculations in 45 minutes for a single collection, versus about 20million x 10 for multiple collections.  Given I'm using 0.98 on correlation threshold, on a single collection, it probably works quite well...not so sure about the re-validation/merging of multiple collections...

Does the re-validation (or merging) using Exp.Gen do the same thing in EAS?  If so, I wonder why the results between those two would be so different ($472 vs $783).

Anyway, just curious on your (and others) thoughts on this.

Thanks