OpenTal - almost new engine

[carldaman]

Great work (and games) mate, can't wait to see how you develop this guy!

Myself, I have 3 new ProDeo personalities and about 5 new Rodents (Kasparov, Positional Vlad, Positional Sergei, Capablanca, etc) still coming, plus a Fruit Reloaded personality.

Too much fun...Alpha who?

Thanks, you got that right, Brendan. Keep up the good work on your end.

Check this thread out in the meantime:

http://www.talkchess.com/forum/viewtopi ... 04&t=66110

[Ovyron]

My goal has been somewhat similar, but I've wanted to get to a point where the machine not only plays like Morphy, Anderssen, Tal, Nezhmetdinov or Shirov, but also can rise above them in strength while retaining the same style, and being (why not?) even stronger than any human alive, past or present.

Great work with Zero, I still believe that there's a way to somehow increase strength without sacrificing style, like some kind of engine that minimizes its own material (doesn't look at material imbalances, but at having the least pieces on the board) while maximizing elo.

I wonder if Shredder's "tripple brain" concept could be used for this:

Have a Master engine running, that loads 6 other engines (say):
Houdini
Komodo
Stockfish
OpenTal
Zero The Hero
Thinker

Then it follows the check:

If OpenTal makes a move that captures a small piece with a big piece (Q x R x N/B x P) or leaves a piece en prise, or sacs a pawn, or puts the king in check, or (add here things that increase style) make that move.

If not, check if Zero The Hero plays such a move, and make it.

If not, check if Thinker makes such a move, and make it.

If not, check if any of Houdini/Komodo/Stockfish makes such a move (unlikely, but who knows), and make it.

If not, check if Tal, Zero, or Thinker, makes a move that any of H/K/S agree with, and make it.

If not, and H/K/S agree on a move, make it (strongest easy move).

If not, you're out of luck - but we'll assume Tal/Zero/Thinker are blundering (i.e. bad move for no reason that increases style), I suggest playing the most original move:

If any of the H/K, K/S or S/H pairs of engines agree on a move, play the one the third one doesn't agree with.

Otherwise, H, K and S suggested a different move.

Make your choice like so:

Check with capture
Check
Capture
Biggest piece (Q R B/N P) moving towards enemy king
Biggest piece moving towards enemy territory
Biggest piece attacking enemy territory

I have lots of B/N in there. As a tie breaker I'd suggest moving a Bishop if you have the Bishop pair, or if you have Bishop and Knight and a queen, but moving a Knight otherwise.

I could play this way manually, though I don't have access to Zero-the-Hero

[carldaman]

My goal has been somewhat similar, but I've wanted to get to a point where the machine not only plays like Morphy, Anderssen, Tal, Nezhmetdinov or Shirov, but also can rise above them in strength while retaining the same style, and being (why not?) even stronger than any human alive, past or present.

Great work with Zero, I still believe that there's a way to somehow increase strength without sacrificing style, like some kind of engine that minimizes its own material (doesn't look at material imbalances, but at having the least pieces on the board) while maximizing elo.

I wonder if Shredder's "tripple brain" concept could be used for this:

Have a Master engine running, that loads 6 other engines (say):
Houdini
Komodo
Stockfish
OpenTal
Zero The Hero
Thinker

Then it follows the check:

If OpenTal makes a move that captures a small piece with a big piece (Q x R x N/B x P) or leaves a piece en prise, or sacs a pawn, or puts the king in check, or (add here things that increase style) make that move.

If not, check if Zero The Hero plays such a move, and make it.

If not, check if Thinker makes such a move, and make it.

If not, check if any of Houdini/Komodo/Stockfish makes such a move (unlikely, but who knows), and make it.

If not, check if Tal, Zero, or Thinker, makes a move that any of H/K/S agree with, and make it.

If not, and H/K/S agree on a move, make it (strongest easy move).

If not, you're out of luck - but we'll assume Tal/Zero/Thinker are blundering (i.e. bad move for no reason that increases style), I suggest playing the most original move:

If any of the H/K, K/S or S/H pairs of engines agree on a move, play the one the third one doesn't agree with.

Otherwise, H, K and S suggested a different move.

Make your choice like so:

Check with capture
Check
Capture
Biggest piece (Q R B/N P) moving towards enemy king
Biggest piece moving towards enemy territory
Biggest piece attacking enemy territory

I have lots of B/N in there. As a tie breaker I'd suggest moving a Bishop if you have the Bishop pair, or if you have Bishop and Knight and a queen, but moving a Knight otherwise.

I could play this way manually, though I don't have access to Zero-the-Hero

I think that constructs such as what you posted have a lot of potential, but they ultimately suffer from a similar weakness that human players have. Namely, they could string many great moves in a row, only to have the next move fail them in a big way, ruining all the prior good work. Such an entity would only be as strong as its weakest link. This is why we haven't seen more dramatic progress along these lines.

What you suggest could be used to plug such holes, especially if the system can be perfected and implemented without significant resource overhead and 'slowdown' effects.

Zero is still a work in progress. Surprisingly, it can quickly find many if not most of the moves AlphaZero played, but tuning it is a real bitch. I hope we reach a point where some processing power can be allocated to a new paradigm, while still retaining the old one to balance things out. That way the whole grail of style and strength working together could be fully attained.

[mclane]

The problems are the material evaluations .
Programs like stockfish / Houdini and Komodo all use the same kind of materialistic evaluation.

Many evaluation features sum up to a value.
Pieces are counted with 1,3,5,9 value no matter if the piece can move on the board or not,
Then a bonus or malus is added to this value.
And the positions are evaluated all in the search tree.

But IMO this is wrong,

It’s not important if the score is x or y.

Important is which side wins or which side has better chances to win.

In the moment you give up this material score, you have a completely different view on chess.

[Andre]

Hello,

Anyone with a Mac who can compile a version?
Thanks in advance!

Andre

[carldaman]

The problems are the material evaluations .
Programs like stockfish / Houdini and Komodo all use the same kind of materialistic evaluation.

Many evaluation features sum up to a value.
Pieces are counted with 1,3,5,9 value no matter if the piece can move on the board or not,
Then a bonus or malus is added to this value.
And the positions are evaluated all in the search tree.

But IMO this is wrong,

It’s not important if the score is x or y.

Important is which side wins or which side has better chances to win.

In the moment you give up this material score, you have a completely different view on chess.

Materialism is a problem, one that can be fixed. We need checks and balances, however. Dynamism is hugely important, but it must be tuned.

[fern]

I just played it. What a dammned beast. I am beginning to hate Norman...

Fern

[fern]

YOU BASTARD!!!!!
What did you do with Rodent?
This Tal thing has defeated me, a strong expert player after all, like a child.

[fern]

Bloated OS?
No sir, a pain in the ass OS
BUT...
at least easy to operate instead of Linux, that ask from you rocket science level

Humble regards
Fern

[PK]

Well, I have played some blitzgames against OpenTal and it is a humbling experience, not necessarily related to engine's strength. Stockfish (or any top engine for that matter) wins because it is objectively much stronger than you. OpenTal wins because you suck at chess. I can see that its play looks suspicious, yet I have no power to prove it.

But perhaps the most uncanny thing is what happens between OpenTal's first and second sacrifice. It parts with a pawn, or with an exchange, and then plays a couple of sane, positional moves, as if nothing of importance happened. And then it sacks again.