but then again who else is crazy enough to test at 20 minutes with a 20 second increment ( equivalent to about 40/53 using the CCRL's crafty benchmark on my hardware)
Version 4.0b would be the released version if released today based on STC results - look for a Holiday release in a few weeks...and yes, the "tactical" setting will be working again in the next release ☺️
went through Ivan Ivec's SF fork called Corchess ( short for correspondence chess ) designed specifically for longer time controls and hand picked the changes ( which was most of them) that worked well in McBrain... stay tuned , more to come...
This is easily one of the best results ever achieved against SF at this time control - but I just cobbled this together - it is Ivan's Corchess code that is making the difference.
The difference between a and b version above was a change of just one line of code - results were in inconclusive with STC. I will add Ivan Ivec to the authors list to the next release of McBrain, it is well deserved!
MikeB wrote:This is easily one of the best results ever achieved against SF at this time control - but I just cobbled this together - it is Ivan's Corchess code that is making the difference.
The difference between a and b version above was a change of just one line of code - results were in inconclusive with STC. I will add Ivan Ivec to the authors list to the next release of McBrain, it is well deserved!
Two of the wins that Stockfish scored above were time forfeits - SF-McB's(versions A) flag fell - I then check my database and found that since August 21st - SF or SF-McB has time forfeited over 60 times ( flag fell ) and it has continued with all the patches since the initial patch on 8/17. - the game above were 20 minute games with 20 second increments and yet the flag fell twice in just under 200 games. It happens less than 1% of the time - but it should almost never happen. McBrain will revert back to the old time code in the next release. The old code worked fine - it was never an issue.
Just checked my test databases from June and November of this year
In June , I ran 55,630 games - Stockfish or SF-McB were involved ~53500 of those games and they had a total of 3 flags or 0.006% or once every ~17833 games or so.
In November so far , I have run 22,431 games - Stockfish or SF-McB were involved in ~18000 of those games and have a total of 28 flags or 0.16% or once every ~643 games . An incident flag rate that is ~27.8 times higher.
MikeB wrote:Just checked my test databases from June and November of this year
In June , I ran 55,630 games - Stockfish or SF-McB were involved ~53500 of those games and they had a total of 3 flags or 0.006% or once every ~17833 games or so.
In November so far , I have run 22,431 games - Stockfish or SF-McB were involved in ~18000 of those games and have a total of 28 flags or 0.16% or once every ~643 games . An incident flag rate that is ~27.8 times higher.
Some what to my surprise , the new time code generates higher ELO than the old timeman.cpp. Apparently the more aggressive use of time earlier in the game more than makes up for the 28x higher Flag incident rate that I experienced. The new time code just has to win one more game out of 643 games to break even and possibly be plus ( often the flag incident is only giving up 1/2 point , turning a draw into loss. I have not seen a game where a win turned into a loss, but I haven't checked every game ). In any event, the new timeman.cpp stays in.
You can hard code an extra buffer to the TM, which will have zero impact at any non-bullet time controls (other than preventing the time losses obviously). The only reason it isn't higher is because SF tests at 10 + 0.10 which gets scaled down even further on fast machines.