1)I did not implement material imbalance table in the evaluation.
2)I use R=3 for null move reduction except not using null move when the remaining depth is 1 because I use different pruning.
3)I do not use double null move but I use verified null move pruning in the endgame as suggested by Omid David.
No threat detection.
4)Nothing special about hash except the fact that 438 is the first version of wbec that is using hash for pruning.
The search of movei is based on normal alpha beta in the way that is done in tscp and I even do not use PVS.
except reductions pruning and extensions and null move
I added to the normal alpha beta fail high and fail low(and the initial aspiration window is not infinite) but I do not have PVS because PVS is also about interior nodes.
I also do not use IID.
The name PVS was misleading and for a long time I thought that the fact that I have a pv means that I use PVS and I even did not care to check what is PVS and later PVS simply did not work for me(maybe because of bugs and I did not do a lot of tries because I prefer to improve the program in other ways).
Uri
Thanks for sharing your approach. Interesting to read... but you didn't mention which change over the last year improved Movei the most. That must be your secret.
A long time ago you described that you didn't do hash pruning and I thought at the time, "uh oh, when you start doing that Movei is going to search even deeper".... and now look at the result.
I tried Omid's NM verify too... but dropped it... might try it again... it seems that every known technique must be tried in every new engine.
As for threat detection.... that worked well in Trace 1.37a but doesn't work at all in the new Trace. Haven't worked out why yet
I'm amazed you don't do IID. I would bet HG Muller would be ABSOLUTELY ASTONISHED too... Movei must have excellent move ordering and/or pruning.
Uri, congratulations.
When do you begun to program Movei?
In the first period that I was a CCC member, 1999-2000,
I don´t remember that you are programing an engine.
1)I did not implement material imbalance table in the evaluation.
2)I use R=3 for null move reduction except not using null move when the remaining depth is 1 because I use different pruning.
3)I do not use double null move but I use verified null move pruning in the endgame as suggested by Omid David.
No threat detection.
4)Nothing special about hash except the fact that 438 is the first version of wbec that is using hash for pruning.
The search of movei is based on normal alpha beta in the way that is done in tscp and I even do not use PVS.
except reductions pruning and extensions and null move
I added to the normal alpha beta fail high and fail low(and the initial aspiration window is not infinite) but I do not have PVS because PVS is also about interior nodes.
I also do not use IID.
The name PVS was misleading and for a long time I thought that the fact that I have a pv means that I use PVS and I even did not care to check what is PVS and later PVS simply did not work for me(maybe because of bugs and I did not do a lot of tries because I prefer to improve the program in other ways).
Uri
Thanks for sharing your approach. Interesting to read... but you didn't mention which change over the last year improved Movei the most. That must be your secret.
A long time ago you described that you didn't do hash pruning and I thought at the time, "uh oh, when you start doing that Movei is going to search even deeper".... and now look at the result.
I tried Omid's NM verify too... but dropped it... might try it again... it seems that every known technique must be tried in every new engine.
As for threat detection.... that worked well in Trace 1.37a but doesn't work at all in the new Trace. Haven't worked out why yet
I'm amazed you don't do IID. I would bet HG Muller would be ABSOLUTELY ASTONISHED too... Movei must have excellent move ordering and/or pruning.
Thanks again, good luck in the Premier class!!!
Ross
I do not know which change is responsible for most of the improvement.
Hash is the main difference between 423 and 438 10 10 10
The difference in ccrl 40/4 is only 40 elo.
I guess that the difference in 40/40 is bigger
