Page 5 of 7

Re: Houdini 5 dev in TCEC

Posted: Fri Sep 09, 2016 8:41 pm
by Henk
JJJ wrote:
Henk wrote:And if these engines may not be corrupt they may lack originality. Don't want to discuss originality. Might be that all these engines are copying (Stockfish) ideas I don't know. Probably normal average growth is fifty ELO points a year.

By the way draw rate of the mutual games of the first four engines is very high. Makes me suspicious too.
The higher the rating, the higher the drawrate. You should know that...
Draw rate of Fire is much bigger than Stockfish.

Re: Houdini 5 dev in TCEC

Posted: Fri Sep 09, 2016 8:43 pm
by JJJ
Henk wrote:
JJJ wrote:
Henk wrote:And if these engines may not be corrupt they may lack originality. Don't want to discuss originality. Might be that all these engines are copying (Stockfish) ideas I don't know. Probably normal average growth is fifty ELO points a year.

By the way draw rate of the mutual games of the first four engines is very high. Makes me suspicious too.
The higher the rating, the higher the drawrate. You should know that...
Draw rate of Fire is much bigger than Stockfish.
Maybe, but it lost a lot against Stockfish. And of course, with more game he would have more loose and more win.

Re: Houdini 5 dev in TCEC

Posted: Fri Sep 09, 2016 8:57 pm
by Dann Corbit
JJJ wrote:
Henk wrote:
JJJ wrote:
Henk wrote:And if these engines may not be corrupt they may lack originality. Don't want to discuss originality. Might be that all these engines are copying (Stockfish) ideas I don't know. Probably normal average growth is fifty ELO points a year.

By the way draw rate of the mutual games of the first four engines is very high. Makes me suspicious too.
The higher the rating, the higher the drawrate. You should know that...
Draw rate of Fire is much bigger than Stockfish.
Maybe, but it lost a lot against Stockfish. And of course, with more game he would have more loose and more win.
As far as taking of ideas, my sig shows my feelings about that.

Back in the day, they had this thing called the ACM, of which I was a member.

People would discover new algorithms and publish them there. The reason was so that other people could use these algorithms and benefit from them.

Not sure what has changed from then to now, but it isn't good.

Re: Houdini 5 dev in TCEC

Posted: Sat Sep 10, 2016 6:35 am
by Lyudmil Tsvetkov
Judging by the few games I have watched Houdini playing in TCEC, at this particular point in time Houdini severely outevaluates both Komodo and SF.

How this was achieved may be a point of speculation, it is possible that search also plays a part in it, but what is certain is that Houdini evals are much more realistic than the evals of the other top 2.

For example, the game Houdini won against Komodo, in the Sicilian out of book H shows 0.76, much more realistic than Komodo's 0.43. It might be the case that the book position is already won for white.

On the other hand, Komodo has something the other 2 top competitors do not have, namely style. Komodo style of play is recognizable, just take knigth outposts if you want, not so Houdini and SF play, I can not recognize any specific style in their play, they just seem to make random moves.

SF's main strength is its fabulous computational abilities in sharp positions.

Judging also by the 2-0 lead in the direct match-up in stage 3, SF should win the superfinal quite convincingly, but it will also depend on the openings used.

I very much hope for short openings, 2-moves neutral book with maybe higher draw rate really produces much more intriguing play to watch than a somewhat biassed 8-move book.

Score is not important for the conoisseurs, I largely prefer 100 interesting fighting draws than the same amount of a bit dullish games with big percentage of convincing book-decided wins...

Anyway, move length is going up...

Re: Houdini 5 dev in TCEC

Posted: Sat Sep 10, 2016 7:18 am
by JJJ
Lyudmil Tsvetkov wrote:Judging by the few games I have watched Houdini playing in TCEC, at this particular point in time Houdini severely outevaluates both Komodo and SF.

How this was achieved may be a point of speculation, it is possible that search also plays a part in it, but what is certain is that Houdini evals are much more realistic than the evals of the other top 2.

For example, the game Houdini won against Komodo, in the Sicilian out of book H shows 0.76, much more realistic than Komodo's 0.43. It might be the case that the book position is already won for white.

On the other hand, Komodo has something the other 2 top competitors do not have, namely style. Komodo style of play is recognizable, just take knigth outposts if you want, not so Houdini and SF play, I can not recognize any specific style in their play, they just seem to make random moves.

SF's main strength is its fabulous computational abilities in sharp positions.

Judging also by the 2-0 lead in the direct match-up in stage 3, SF should win the superfinal quite convincingly, but it will also depend on the openings used.

I very much hope for short openings, 2-moves neutral book with maybe higher draw rate really produces much more intriguing play to watch than a somewhat biassed 8-move book.

Score is not important for the conoisseurs, I largely prefer 100 interesting fighting draws than the same amount of a bit dullish games with big percentage of convincing book-decided wins...

Anyway, move length is going up...
Did you like how Houdini played the queen gambit for its first game on rapid TC ?

Re: Houdini 5 dev in TCEC

Posted: Sat Sep 10, 2016 10:28 am
by Lyudmil Tsvetkov
JJJ wrote:
Lyudmil Tsvetkov wrote:Judging by the few games I have watched Houdini playing in TCEC, at this particular point in time Houdini severely outevaluates both Komodo and SF.

How this was achieved may be a point of speculation, it is possible that search also plays a part in it, but what is certain is that Houdini evals are much more realistic than the evals of the other top 2.

For example, the game Houdini won against Komodo, in the Sicilian out of book H shows 0.76, much more realistic than Komodo's 0.43. It might be the case that the book position is already won for white.

On the other hand, Komodo has something the other 2 top competitors do not have, namely style. Komodo style of play is recognizable, just take knigth outposts if you want, not so Houdini and SF play, I can not recognize any specific style in their play, they just seem to make random moves.

SF's main strength is its fabulous computational abilities in sharp positions.

Judging also by the 2-0 lead in the direct match-up in stage 3, SF should win the superfinal quite convincingly, but it will also depend on the openings used.

I very much hope for short openings, 2-moves neutral book with maybe higher draw rate really produces much more intriguing play to watch than a somewhat biassed 8-move book.

Score is not important for the conoisseurs, I largely prefer 100 interesting fighting draws than the same amount of a bit dullish games with big percentage of convincing book-decided wins...

Anyway, move length is going up...
Did you like how Houdini played the queen gambit for its first game on rapid TC ?
Quickly browsed through the game.
This is rapid.
Did not strike me with anything in particular.
I consider the QG very favourable for white.
Indeed, where did Fruit make a fatal mistake? It just lost without making any obvious mistake.

Any guess as to when the superfinal will start?

Re: Houdini 5 dev in TCEC

Posted: Sat Sep 10, 2016 10:51 am
by beram
Lyudmil Tsvetkov wrote:
JJJ wrote:
Lyudmil Tsvetkov wrote:Judging by the few games I have watched Houdini playing in TCEC, at this particular point in time Houdini severely outevaluates both Komodo and SF.

How this was achieved may be a point of speculation, it is possible that search also plays a part in it, but what is certain is that Houdini evals are much more realistic than the evals of the other top 2.

For example, the game Houdini won against Komodo, in the Sicilian out of book H shows 0.76, much more realistic than Komodo's 0.43. It might be the case that the book position is already won for white.

On the other hand, Komodo has something the other 2 top competitors do not have, namely style. Komodo style of play is recognizable, just take knigth outposts if you want, not so Houdini and SF play, I can not recognize any specific style in their play, they just seem to make random moves.

SF's main strength is its fabulous computational abilities in sharp positions.

Judging also by the 2-0 lead in the direct match-up in stage 3, SF should win the superfinal quite convincingly, but it will also depend on the openings used.

I very much hope for short openings, 2-moves neutral book with maybe higher draw rate really produces much more intriguing play to watch than a somewhat biassed 8-move book.

Score is not important for the conoisseurs, I largely prefer 100 interesting fighting draws than the same amount of a bit dullish games with big percentage of convincing book-decided wins...

Anyway, move length is going up...
Did you like how Houdini played the queen gambit for its first game on rapid TC ?
Quickly browsed through the game.
This is rapid.
Did not strike me with anything in particular.
I consider the QG very favourable for white.
Indeed, where did Fruit make a fatal mistake? It just lost without making any obvious mistake.

Any guess as to when the superfinal will start?
I guess Rapid TCEC will take about max 6 six weeks

and The first mistake was probably Bf5? after move 17 Qa4 - where Rb8 (as shown in the PV by Houdini) was the much better option
After 17 - Rb8 18 h5 doesnt work as good as after the played Bf5

[d] r1bbr1k1/1p3ppp/p1n3q1/8/Q6P/P1N1PN2/1P3PP1/2R1KB1R b K - 0 17

Re: Houdini 5 dev in TCEC

Posted: Sun Sep 11, 2016 12:10 pm
by Lyudmil Tsvetkov
I did the effort to check most of the Houdini games of stage 3 of TCEC, and can draw the following conclusions:

- at this point in time H has by far the best eval, much better than Komodo, much better than SF

- Houdini outposts are greatly improved, now on a par with Komodo outposts(never had that in the past, at least to the level of Komodo)

- H already knows about bad bishops, pawns on squares the colour of the bishop(never had that in the past)

- H has very good understanding of connected pawns, probably also chains to some extent, better than both Komodo and SF understanding

- significantly improved eval of weak pawns(never had that in the past)

- excellent, absolutely wonderful treatment of pawn shelter, fully the way I want to see it implemented; neither Komodo, nor SF are anywhere close to this, on the contrary, they are quite conservative in their pawn shelters, often leaving pawns on their initial squares, while H pushes them wildly, leaving its king bare; of course, when you tune your pawn shelter to the rigth terms like chains, connected pawns, etc., you can play with almost no shelter, storming the enemy position instead(never had that in the past)

- main H problems are visible in the search, where it is regularly outplayed, at least by SF, but also strangely misses continuations in some promising positions

So, no matter what Robert will claim, most of the improvements are due to greatly improved eval function.

Conclusion is, if someone wants to check rigth position assessment, go with soon-to-be-released H 5, H assessment in most positions will be much better than both Komodo and SF.

Re: Houdini 5 dev in TCEC

Posted: Sun Sep 11, 2016 12:18 pm
by Lyudmil Tsvetkov
I forgot to say that H still keeps its good attacking style from the past, above features only enriching it, while making it significantly more positional

PS. Is not it possible to delete that very-stupid benchmark-new-smartphones thread? It has already reached 2000 posts, fully unreadable, dating from a couple of years, fully unrelated to the forum content, and, most importantly, always appearing/being renewed(maybe on purpose) to make difficult normally reading other threads???

Re: Houdini 5 dev in TCEC

Posted: Sun Sep 11, 2016 12:45 pm
by brianr
I for one find the handheld device benchmarks quite interesting (and highly relevant). While it has many posts, an updated table is periodically included so as to not have to go too far back. Moreover, there is a lot of work behind those posts.
Lyudmil Tsvetkov wrote: PS. Is not it possible to delete that very-stupid benchmark-new-smartphones thread? It has already reached 2000 posts, fully unreadable, dating from a couple of years, fully unrelated to the forum content, and, most importantly, always appearing/being renewed(maybe on purpose) to make difficult normally reading other threads???