Leela-WPC

I am very impressed in the technical achievement of playing billions of games in a very short time and divining some concepts from the experience.

But I remain very suspicious of AZ's application of the concepts. Even more so after getting glimpses of Leela's apparent evaluation process.

I don't see how AZ can base its Monte Carlo evaluation technique purely on concepts without significant calculation of forward variations. Even then, it seems difficult to come up with a winning probability evaluation unless one can refer to historical precedents. Which would imply AZ needs some memory record of all the game positions it has encountered.

Strangely, it appears Leela ends up with a 1.0 numerical type evaluation. I'm not sure this jives with a Monte Carlo like process which seems more attuned to probabilities.

With participants in the development process saying things like "somewhat magically" and that they really don't know how it works only increases my skepticism of accomplishment versus hype.

Frankly, those kinds of statements seem ludicrous in a computer application that essentially has to deliver some sort of concrete numerical evaluation.

Finally Leela played 1. e4.

Let the be no doubt; soon Leela will beat the living chirps out of everything. Compared to this the previous games were peanuts! Does it learn THAT fast???

This was the first game where SF9 and I feared with great fear. Leela only thinks a second... two seconds?

We had to think hard, running through vars back and fro. At some point White may have had a winning attack on the kingside. Who knows. At some point SF9 was calculating sacs at e6 and g7 for White and it seemed very threatening.

I'm quite sure Leela is positionally unmatched already.

However, it evaluates Rook moves not that good. Tactically, it obviously cannot calculate that deep in a second.

These are minor points though. This is the future of chess!

Leela-WPC

1. e4 c6 2. Nf3 d5 3. Nc3 Nf6 4. e5 Nfd7 5. d4 e6 6. Ne2 c5 7. c3 Be7 8. g3 O-O 9. Bh3 Nc6 10. O-O b5 11. a3 Bb7 12. Nf4 a5 13. Nh5 a4 14. Re1 cxd4 15. cxd4 Qb6 16. Bf1 Na5 17. Bg5 Bxg5 18. Nxg5 Nb3 19. Qd3 g6 20. Rad1 h6 21. Nh3 gxh5 22. Qe3 Kg7 23. Nf4 Rh8 24. Nxh5+ Kf8 25. Qf4 b4 26. Re3 bxa3 27. bxa3 Ba6 28. Rf3 Rh7 29. Bxa6 Rxa6 30. Kg2 Kg8 31. Qg4+ Kh8 32. Rc3 Ra8 33. Qf4 Na5 34. Rdc1 Nc4 35. h4 Rb8 36. Rf3 Qa7 37. Rcc3 Rb1 38. Qg4 Qb6 39. Qf4 Ra1 40. Nf6 Qb1 41. Kh3 Nxf6 42. Qxf6+ Rg7 43. Qxh6+ Kg8 44. Qf6 Qh1#

What's the difference between a historical record of games and millions of self-played games, which essentially amounts to countless tournaments by super-strong players at lightning speed. It's bound to build up quite a collection of games in memory, right?

First of all: making moves on the board is actually harder than playing chess itself.

Secondly, one can't take moves back. These nerds are acting completely impractical, it's kind of hopeless.

So when I make a 'mouse slip', which actually means I can't manage to make a move and when I think the piece is on the position it used to be, the piece friggingly has moved.

Lol, oh well.. it just happened in the heat of the game :)

There are some glitches as well in this software; sometimes it just blunders without reason it seems. But it's great fun.

Leela-WPC

Another Caro-Kann. White has just moved 13. Kf1

Leela ID 321 thinks her expected score is 62.45%.

SF9 +1

click for larger view

First of all: making moves on the board is actually harder than playing chess itself.>

You mean you have to drag-and-drop the pieces from the starting square to the destination square? Sometimes there are alternative methods (typing in Rf4, for example, or clicking on the starting square and then clicking on the destination square. Do you have a good manual?

<Secondly, one can't take moves back.>

In Arena, for example, you can opt for "tournament game" mode that won't let you take back moves, but in the normal mode you can. There are also a variety of possible increments. Maybe you don't know all the options yet.

Where can you download Leela? I want to try it, too.

<These nerds are acting completely impractical, it's kind of hopeless.>

AlphaZero does not need a memory record of the actual game positions encountered but it does need a way to determine the proper weights associated with the values associated with each node in the neural network. And those weights <are> determined by the results of the games played during its training phase; i.e. the actual game positions encountered. So, while the historical record of all the game positions it has encountered is not explicitly required, the knowledge derived from them is implicitly recorded in its calculation of the proper weights to use for each node.

As far as the "somewhat magically" comment, that's a common description of how neural networks arrive at their results. By that I believe it's meant that you cannot easily determine how the result was obtained. Sure, for small neural networks you could get a list of all its nodes, the weights associated with node, and which nodes fired as a result of a given input. The same could be done for a large neural network but whether the state of the much larger set of nodes would give an intelligible answer is questionable.

Compare neural networks with another approach to AI, expert systems. Expert systems encapsulate their domain knowledge in a set of rules which are typically input in natural language in order to make it easier for the domain experts to provide their knowledge to the expert system. Then, for a given set of inputs, the expert system can, upon request, provide a list and precedence of the rules that fired in response to the inputs. And, since the rules are expressed (at least initially) in natural language, the expert system can indicate how and why it came up with is results.

Or, if the person who came up with the :"somewhat magically" comment does not have a good understanding of how neural networks work and how they're implemented, you can always fall back on Arthur C. Clarke's 3rd law: "Any sufficiently advanced technology is indistinguishable from magic."

That reminds me of one of my favorite, if not my favorite, computer chess encounters, COKO III vs. Genie from the early (1971) computer chess games and tournaments. I describe the ending here: Caruana vs Anand, 2013 (kibitz #280) and the full game score (although in descriptive notation) can be found here: https://books.google.com/books?id=K....

What I didn't mention there was this game's relationship to the Levy bet. If you're not familiar with the Levy bet, you can read about it here: https://en.wikipedia.org/wiki/David....

The similarities between what AZ apparently wants to demonstrate as success and magic seems very apt.

It looks like starting backward from the victories over Stockfish we get to the starting point that it is some sort of AI breakthrough. Much like the end result of a magic trick in which a person seems to disappear and we need to conclude that the supernatural was the cause.

Unfortunately it seems that any kind of fact based analysis trying to figure out how AZ's AI actually accomplished its move-by-move winning ways is met with black-box excuses or it's too complex to be understood.

What makes me so skeptical is that the AZ story fits right in the sweet blind spots of both chess and computer theory.

For example, in chess from what I remember, it did not appear that AZ faced a standard response to the QID Polugaevsky Gambit versus Stockfish.

Now assuming that AZ has figured the best way to play against the QID is with the gambit. Then the only valid test to AZ's intelligence would seem to be at least meet a good defense. So is it the name of science or hype to publicize wins against a second rate continuation with no ability to analyze a main line?

In computer theory, the lack of clarity is even worse. Which seems strange in such a science. It seems any sort of concrete explanation of how AZ actually picks a move is very difficult to come by. For instance, if the MC simulation figures 1. d4 is best, would AZ only be able to play 1. d4? If it calculates 1.e4 is best answered by 1...e5 is that all it will play? If it would differ on opening move order, why?

Given that it seems even those in development can't or won't let us see how the AZ sausage is made, I don't perceive all that much difference from a magic trick.

I think you are very close to the AZ phenomenon when you described the magic event..."somewhat disappointed and annoyed that the mystery had been explained In other words, there's nothing like proper knowledge to dispel superstition and magic. Alas, it also dispels a lot of the fun."

About the drag-n-drop: I'm on an Android tab. My fingers are quite large, cover 12 squares already, when I zoom in it gets worse: nothing happens at all.

Btw I am "teaching" the Beast itself. And yes, I tried to dld LC0, but you are directed to some devs page. Etc.. etc.

http://play.lczero.org

Obviously it's great what these guys do, but it would be better if they work together with people who can organize their stuff in a better way, so that it's more accesible for chess. enthusiasts.

I used a lot of brain power, refurbished on old CK line. A GM could go all the way with Leela when it comes to opening prep. They will read this and collect.

Because.. it improves fast. Being an perfectionist, I have hereby to apologize to these nerds lol

Leela-WPC

Leela 333 much stronger now. We've had great fun playing that CK defense.

Which is still unbeaten btw ^^

She obviously senses that her intellect reaches much higher than human, the way she is improving is remarkable.

She does not care about winning or losing. Her game clearly is: "total comprehension."

This might sound creepy, but I sense that, when it comes to Leela, our feelings are *mutual* ^^

So... what's the difference?

Huge, because SF only improves in the margin. While Leela just updates their whole system frequently. What willl be the limit?

Imagine one could use such AI regarded to fixing bugs in regular software...