McBrain 2017 2.4 Planned Release Date May 26, 2017

[MikeB]

I accelerated this release so that you can match McBrain ( a Stockfish engine) against the latest Komodo which is coming out this Monday , May 25th, over the upcoming Memorial Day Holiday.

Play by ELO was totally revamped.

For those of you who used the play by ELO feature, ELO was set to fixed number of nodes - period - no matter what the time control.

This function was totally re-wriiten and now it will target an nps based on the ELO level selected and the toal nodes search will increase at the the target nps rate. The range of Elo was rescaled from 1725 to 2825. The target nps at ELO 1725 is 32. The target NPS at 2825 is 65536 roughly in line with a high end Revelation machine using Stockfish. The function the accepted motion that ELO increases 100 Elo for every doubling. Not surprising that's a pretty generic assumption and it does not accurately reflect how Stockfish scales with increases in nps.

Stockfish gains in ELO by increasing nps are not linear at these lower depths. There is a sweet spot between 8000 NPS to 65000 NPS where Stockfish increases ELO at rate greater than 100 ELO per doubling NPS.

What's the point - and you may be wondering how I get to ELO of 2825 with 65000 NPS. The short answer is that I needed to keep it simple.

I began with 65000 NPS because I wanted the play be ELO function to be near the playing level of the $3500 dedicated Revelation Chess computer , and with Stockfish I have read Revelation gets over 50000 NPS. I then then ran thousands of games against Toga 3 ( about 2880 CCRL 4/4) single core) and came up with an ELO over 2900 - let's call it 2950. And that is pretty much in line with the people who would know more than me say about Revelation's ELO. So far so good. I then ran the engine with ELO set to 2600 ( based on max setting of 2950) against the same Toga engine - I had a surprise - those results were no where near 2650 - the rating came in under 400 points below Toga. To keep it simple , I split the difference between the two ratings and came up with the 2825 as the high end and everything below that is calculated on the doubling equals 100 ELO notion. . The 2600 level may be overstated 125 ELO when matched against Toga - the point is - this is not an exact science - and in the end it does not really matter, as these are just reference points you use to mark you personal progression ( or regression).

Code: Select all

SF NPS		Rating
		
   32		 1,725
   64		 1,825
  128		 1,925
  256		 2,025
  512		 2,125
 1,024		2,225
 2,048		2,325
 4,096		2,425
 8,192		2,525
16,384		2,625
32,768		2,725
65,536		2,825

Any ELO between 1725 and 2825 can be selected ( over 1000 levels). The program uses a random function that will will select an ELO level at random that will be within 33 ELO of the ELO selected. The engine will determine the nodes to search , based on the ELO selected and the calculated time it plans to spend on that move. Say one selects ELO 1825 and you are playing ~10 second per move, in total the engine will look at 640 nodes (10*64). It has built non-busy sleep function to delay it 's response until the appropriate time is up and that also factors in how may seconds ( approximately) it will take to analyze the total number of nodes it will search ( only matters at the higher levels). There will be no delay while engine is in book ( this is also different from before from old version the using Move_Delay - move delay is now automatic when using play by ELO.

And of course it will include all the features as before and all the great code updates by the SF Development team - the SF team had a great week - in two days, 3 green patches totaling over a 10 ELO gain passed.

Also, thank you all for your support , it is much appreciated. This week McBrain reached over 1700 downloads in total.

Current Release:
https://github.com/MichaelB7/Stockfish/ ... s/tag/v2.3

Version 2.4 with the revamp Play By Elo feature will be released on or about Friday May 26th, 2017. Notice will be posted here on Talkchess.com