New Houdini

[shrapnel]

Hi Dann, thanks for your support.

I'm working on the Tactical Mode right now.
It's not easy to reach the same level as the Tactical mode of Houdini 3 and Houdini 4. Houdini 3 is relatively the strongest tactician (relatively meaning "compared to its strength"). After Houdini 3 the strength improvements have actually made the engine less tactical.

I'm trying to make the "Tactical Mode" compatible with normal search, so that you can run, say, 5 threads with normal search combined with 1 thread in Tactical Mode.
Maybe I can get this operational for the TCEC superfinals .

Cheers,
Robert

WOW ! A rare Post from the Great Man himself !

[Laskos]

From this TCEC Rapid, to get one more glimpse about Houdini, I plotted the distribution of won games as length of the games (thanks to Ferdy for the tool to extract data). It gives us additional information compared to simple score at this time, which stands at 35/38 Houdini, 32.5/38 Stockfish, 31.5/37 Komodo.



Average won game length: Houdini 51.6 moves, Stockfish 50.9 moves, Komodo 55.1 moves. From the plot, it seems Houdini grinds its opponents in the middlegame. Both Komodo and Stockfish seem to take it into endgame more often (Komodo even more than SF). People following TCEC Rapid close up probably have clearer impressions. All in all, the score doesn't seem a fluke, Houdini seems competitive.

[Dirt]

WOW ! A rare Post from the Great Man himself !

Yes, it's very good to hear from him.

[Lyudmil Tsvetkov]

I've been watching some of Houdini's TCEC games. They are impressive. The engine seems to have a punchy agressive style too. I wouldn't be surprised if it isn't close to parity with Stockfish.

Steve

Your impressions seem to be confirmed as of now. Not only Houdini leads and has most wins, its wins are on average faster (less moves):

Code: Select all

         SF    Hou    Ko
         25    30     26    <-- number of wins  
Length            
21-30     1     0      2
31-40     5     7      1
41-50     8     9     10 
51-60     2     9      5
61-70     6     2      3
71-80     2     2      1
81-100    1     1      4
=========================
Won      52   50.5    55.4
Game
Length

I do not follow TCEC, can it not be that Houdini's opponents resign earlier?

No, simply Houdini plays better chess, a bit better than SF and Komodo, but nothing earth-shaking.

I bet after version 5 release H will top most rating lists 20-30 elo above SF 8 (which should be released about the same time) with contempt set on, about fully equal with no contempt.

I do not know about TCEC final though, everything will depend on the openings used, with shorter openings and more closed ones I would favour slightly Houdini, If the 8-move GM suite is used, which would be a major mistake to me, as will only bring more draws because of exhausted positions and less interesting play, especially if there are a lot of Sicilians and complex open positions, SF will undoubtedly be on top.

In any case we are heading into a very interesting final.

[Werewolf]

No, simply Houdini plays better chess, a bit better than SF and Komodo, but nothing earth-shaking.

I bet after version 5 release H will top most rating lists 20-30 elo above SF 8 (which should be released about the same time) with contempt set on, about fully equal with no contempt.

OK, personally I doubt that. The game sample seems too small to draw this conclusion. Also the official release SF compiles are sometimes a bit faster, adding extra elo.

[Ozymandias]

I'm working on the Tactical Mode right now.
It's not easy to reach the same level as the Tactical mode of Houdini 3 and Houdini 4. Houdini 3 is relatively the strongest tactician (relatively meaning "compared to its strength"). After Houdini 3 the strength improvements have actually made the engine less tactical.

I'm trying to make the "Tactical Mode" compatible with normal search, so that you can run, say, 5 threads with normal search combined with 1 thread in Tactical Mode.

I've seen some tactical test suites support that claim. H4T isn't stronger than H3T, and it's tactically less savvy. Maybe the problem was in trying to keep it as close as possible (strength wise) to the default engine. As long as H5T turns out to be stronger than its predecessors, I personally don't care if it falls 50 ELO point below default, as long as it improves H3T where it matters.

That's an option full of possibilities, if the user has control over how many cores are assigned to each personality. Further control over the tactical mode, beyond the mere 0,1 and 2 settings, would only multiply those possibilities.

[Jouni]

In my tests SF Matefinder 110416 is slightly better than H3 tactical. But is it anymore updated, http://chess.ultimaiq.net/stockfish.html version is crap?!

[Lyudmil Tsvetkov]

No, simply Houdini plays better chess, a bit better than SF and Komodo, but nothing earth-shaking.

I bet after version 5 release H will top most rating lists 20-30 elo above SF 8 (which should be released about the same time) with contempt set on, about fully equal with no contempt.

OK, personally I doubt that. The game sample seems too small to draw this conclusion. Also the official release SF compiles are sometimes a bit faster, adding extra elo.

I am not looking at the game sample, but at the way the engines play.
Sometimes, a single game is enough to tell the difference.

I fully do not understand those compile issues. Certainly Robert is able to make faster or slower compiles, just as Komodo and SF are able to. We talk software, not compiles and assembler versions.

[Leo]

What are Babson tasks?

[Dann Corbit]

https://en.wikipedia.org/wiki/Babson_task