Here's Hoping They Don't Break Chess Next...

[Vadim Tsypin]

Re: Here's Hoping They Don't Break Chess Next...

This was predicted (I'm sure in other places as well) in the Ender's Game universe where an AI takes over managing Ender's money and almost immediately does a better job of it than the humans could have just by examining past behaviour...

Chris, I took a few days to get and read Ender's Game (the original book) since I heard about it before but never actually held it in my hands.

Your comment makes me think: the siren song of AI appears strong. Driverless cars, applications like Hopper that tell me "Don't purchase this ticket now!" :), doctor-less diagnoses in fields like radiology and dermatology, now a robo-advisor that would manage my money better that I'd ever be capable to... - the world becomes so safe and secure for us, humans, doesn't it? Maybe The Singularity will result not in anti-utopia but in a genuine utopia where humans would be:

• relieved of mundane tasks and free to concentrate on creative pursuits ("our" prospective)
• happy and safe as toddlers playing in a secured sandbox (AI's prospective)

?

[Paul Bonham]

Re: Here's Hoping They Don't Break Chess Next...

Lol "Weasel words" aka Adverbs.

Anyway... Chess basically can't be mathematically solved, but if a computer were to be developed that could beat every other computer on a consistent basis while drawing against an identical version of itself 100% of the time, we've "effectively" solved the game.

I know insulting people is a past time for many chesstalkers, but I'm making a good point here. Computers are already taking some of the 'magic' out of chess, but a self learning AI that supersedes any existing calculation based software would be a huge blow.

These ARE good points, Matthew, and I agree with them. You even mention the "magic" being removed from chess, which I'm a bit surprised Ed Seedhouse doesn't recognize. I'm sure Ed must remember the times before Houdini and Stockfish, when there was "magic" in a match between Kasparov and Deep Blue. The reason there was "magic" is that in those times, analysis of chess was something the elite of chess could make a living from -- they could write books analyzing openings, middlegames, endings and there were no computer engines to have the final say. There was no final say, there was only the opinion of all the elite of chess. Computers have basically removed that aspect of chess. Computer engines today have the final say in all chess analysis except the most contrived ones designed to frustrate their horizon effect.

I believe that the effect of computer engines will become more pronounced in the very near future. Chess is getting old but it is not that easily defeated. It IS effectively solved just as you describe, nevermind Ed's protestations. Just look at correspondence chess at its highest levels (where players use engines as well as their own abilities): about 90% draw rate. And it is dying out rapidly.

But.... as Darwin would say, evolve or perish. Chess must evolve if it is to survive. And again, Ed's statement that human foot races have survived despite motorcycles isn't really relevant. First of all, human foot races don't exactly set the world afire. They are entertainment, like the theater. The competition aspect only adds to the entertainment value -- if each runner was running by himself and the fastest time won, far less people would watch. Chess by comparison has a far, far reduced entertainment value (almost nil in fact to people who themselves don't play chess). Even now, we know that Magnus Carlsen is not World Champion, and just as we would call Usain Bolt "World's Fastest Human" we should call Carlsen "World's Best Human Chess Player".

There are ways to do the 2 things that are demanded: make chess resistant to computer engines, and increase the entertainment value (even to non-players). This will be coming out in the near future and chess skills will provide new avenues to livelihood. Chess is a living thing, and all living things are subject to evolution.

[Ed Seedhouse]

Re: Here's Hoping They Don't Break Chess Next...

You even mention the "magic" being removed from chess, which I'm a bit surprised Ed Seedhouse doesn't recognize.

Yeah, sorry, don't believe in "magic". Chess is a complex war game and the idea is to win while complying with the rules. Sometimes, if the two players are right, a game may be "artistic". We take aesthetic pleasure from playing such games over, but the loser seldom gets a share of the admiration for making such a game possible. As long as human beings compete with each other such games will be produced from time to time.

But if you want "magic" it's easier to just watch an earthrise (commonly called "sunset"). Of course that isn't actually "magic", it is merely "wonderful".

I see no reason to change chess just because computers now exceed our ability at it. Anything we do to make chess difficult for computers will, I think, make it "not chess". I still like to play chess against people and my computer makes that possible now that my physical limitations pretty well prevent me from playing it in person. In that sense computers have been a great boon to chess.

I think chess can be made a better game with a slight change in the scoring rules, but not by trying to make it intractable for computers and tractable for humans. Doing that will, I suspect, either make the rules absurdly complex or the board vastly bigger with new and weird pieces. If you scale "go" up to, say, 25x25 you still have "go". Same with "draughts" and some other games (reversi?). Any attempt to scale chess up will, I think, result in it not being chess any more.

And I also think that if you believe computers have taken the "magic" out of chess then your understanding of chess (as opposed to your strength at playing it) is impoverished.

[Garland Best]

Re: Here's Hoping They Don't Break Chess Next...

... make chess resistant to computer engines ...

Here is where I now question if it is possible. Can one make a game of intellect that a player can consistently outperform a computer? Based on the efforts of IBM (Watson) and Google (AlphaGo Zero), I seriously doubt it.

[Matthew Nicholson]

Re: Here's Hoping They Don't Break Chess Next...

I believe that the effect of computer engines will become more pronounced in the very near future. Chess is getting old but it is not that easily defeated. It IS effectively solved just as you describe, nevermind Ed's protestations. Just look at correspondence chess at its highest levels (where players use engines as well as their own abilities): about 90% draw rate. And it is dying out rapidly.

I remember hearing something about a 95%+ draw rate in the Berlin at CCGM level. Seeing as my chess understanding is impoverished as a 2300+ and climbing player, I study these correspondence games quite a lot as part of my opening prep. What's surprising me is that in these Berlin/Slav/SemiSlav/QGD/Petroff games, white players aren't even getting obvious 'pressure'. The games just seem equal from start to finish. In my eyes correspondence chess is very nearly dead. No magic in a prepared draw.

[Paul Bonham]

Re: Here's Hoping They Don't Break Chess Next...

Here is where I now question if it is possible. Can one make a game of intellect that a player can consistently outperform a computer? Based on the efforts of IBM (Watson) and Google (AlphaGo Zero), I seriously doubt it.

I don't know anything about the game of Go, nor about what AlphaGo Zero is doing. It could turn out that AlphaGo Zero has found a "flaw" in the game of Go that makes it not as computationally difficult as humans have thought until now. We hear a lot about data mining these days: computer algorithms that find patterns in data where the set of data is so huge that humans can't even manage to look for patterns, let alone see them. So on a 19 x 19 Go board, humans can only get a tiny grasp of all the possible patterns of black and white stones. Computers can do a much better job of searching through terabytes, or even petabytes, of board pattern data, and it wouldn't surprise me if AlphaGo has managed to find something meaningful that is escaping all the human players.

Getting back to chess, I do strongly believe there is a way to go beyond the simple model that computer chess engines use (minimax search algorithm with alpha-beta pruning). The key is to present them with something they can't evaluate.

What is an example of something that can't be evaluated? There are plenty of real world examples. The stock market provides a good case study. Both humans and computers can beat the market, but neither can achieve perfection or anything close to perfection in investment choices. Only by backtracking through historical data can one determine the exact points in time when it would have been ideal to buy and to sell.

So, what if in chess there was something you could do that was generally good but could only be done a certain number of times? This was the idea behind my variant of Option Chess, and the "something" that gets added is the ability to make 2 moves on your turn (with important limitations, such as only the second move can be either a capture or a check). This is called an "option" and each player only gets 12 of them to use between moves 9 and 32 inclusive. Then at move 33, each player gets 4 more options.

How would an engine decide whether to use an option now versus saving it for later? The chess engine would be limited to doing the same thing the human player is doing: using heuristics (rules of thumb). One of these heuristics is that options become more valuable (lead to greater gains) as pieces disappear from the board. But the chess engine, just like the human, cannot see deep enough to know just how valuable an extra option will be 30 moves in the future. Save too many options for the endgame and you could lose the game before the endgame even arrives. So another heuristic would be to set a positional evaluation limit, and if that limit gets breached, then go ahead and use an option now instead of saving it.

In effect, humans and computers would each be using their own heuristics. But humans can adjust heuristics, whereas the engine's heuristics are mostly built in. Programmers would have to learn to make their heuristics as flexible as possible, but even knowing when and how to do that ALSO involves heuristics!

I truly believe this would result in equality between humans and engines in Option Chess. But only time will tell. The first thing needed is for humans to get good at Option Chess and for programmers to want to write Option Chess engines.

[Sid Belzberg]

Re: Here's Hoping They Don't Break Chess Next...

whereas the engine's heuristics are mostly built in.

The whole point of Alpha Go Zero is that thing learns tabula rasa. So I am confused by this assumption. Here is an interesting project on Github which is a version of Alpha Go Zero in C++ that is written for a public distributed computing network that they are working on setting up. The current version Alpha Go Zero works only on Google's proprietary Tensor Processing Units.
At first glance the code here appears to be amateurishly put together by a hobbyist with things like "If Else" statements instead of simply case statements and so on. But a very cool effort none the less!
https://github.com/gcp/leela-zero

[Paul Bonham]

Re: Here's Hoping They Don't Break Chess Next...

The whole point of Alpha Go Zero is that thing learns tabula rasa. So I am confused by this assumption. Here is an interesting project on Github which is a version of Alpha Go Zero in C++ that is written for a public distributed computing network that they are working on setting up. The current version Alpha Go Zero works only on Google's proprietary Tensor Processing Units.
At first glance the code here appears to be amateurishly put together by a hobbyist with things like "If Else" statements instead of simply case statements and so on. But a very cool effort none the less!
https://github.com/gcp/leela-zero

Sid, I wasn't referring to AlphaGo Zero in referring to heuristics, the context was a chess engine (any chess engine) adapted to play Option Chess. Now, if a neural net chess engine were trained as you say tabula rasa (Latin for blank slate) to play Option Chess, it still would have no better method to use (when deciding when to play an option) than heuristics, exactly as a human would do. And what, if anything, would make the computer's heuristics better than those of all humans? That is the 64 million dollar question (inflation).

Thanks for the link, I will have to check that out. Have you determined anything by looking at the github code, such as whether the engine is doing some kind of data mining during its training?

[Sid Belzberg]

Re: Here's Hoping They Don't Break Chess Next...

Sid, I wasn't referring to AlphaGo Zero in referring to heuristics, the context was a chess engine (any chess engine) adapted to play Option Chess. Now, if a neural net chess engine were trained as you say tabula rasa (Latin for blank slate) to play Option Chess, it still would have no better method to use (when deciding when to play an option) than heuristics, exactly as a human would do. And what, if anything, would make the computer's heuristics better than those of all humans? That is the 64 million dollar question (inflation).

Thanks for the link, I will have to check that out. Have you determined anything by looking at the github code, such as whether the engine is doing some kind of data mining during its training?

Have you determined anything by looking at the github code, such as whether the engine is doing some kind of data mining during its training?

I only looked at the code superficially but yes it appears to use a Monte Carlo Tree Search and a deep residual convolutional neural network stack that collects the probabilities of various positions occurring and establishes hierarchical classifiers. Your option Chess has exponentially increased the possibilities however the number of possible go games is magnitudes larger then the number of possible traditional chess games so I am not sure if your statement about "no better method then heuristics" is correct as it appears to me that heuristics are teased out within hierarchical classifiers which is not much different then how the human brain works. So far computers appear to be far more effective at this then humans under certain circumstances such as board games.

[Ed Seedhouse]

Re: Here's Hoping They Don't Break Chess Next...

I don't know anything about the game of Go, nor about what AlphaGo Zero is doing.

And next to nothing about computation, judging from what follows...

[Paul Bonham]

Re: Here's Hoping They Don't Break Chess Next...

I only looked at the code superficially but yes it appears to use a Monte Carlo Tree Search and a deep residual convolutional neural network stack that collects the probabilities of various positions occurring and establishes hierarchical classifiers. Your option Chess has exponentially increased the possibilities however the number of possible go games is magnitudes larger then the number of possible traditional chess games so I am not sure if your statement about "no better method then heuristics" is correct as it appears to me that heuristics are teased out within hierarchical classifiers which is not much different then how the human brain works. So far computers appear to be far more effective at this then humans under certain circumstances such as board games.

Agreed, not much different from how the human brain works, that was the point I was making. Now, the board games are where the engines shine because the board games all lend themselves to an eval() function in which 100% of the information that is needed is present and accounted for. The difference with Option Chess will be that an option saved for 20 or 30 moves from now cannot be evaluated (accurately), it can only be estimated much as stock prices and movements are estimated. It isn't about increasing the possibilities, it's about not knowing the possibilities accurately and oftentimes being WAY off. I hope you do realize this is a new concept that neither chess nor go encapsulate.

An even better example: the draft choices that football teams make from among a pool of college players. How many times do those choices turn out to be nothing like what was expected? More often than not is the truthful answer. A computer would do no better than humans at making those draft choices.

I'm not saying the engines won't play well, I'm saying they won't dominate humans the way current engines do in standard chess.

But yes, this has yet to be proven in the case of Option Chess. I'm quite confident that humans can and will match future computer engines.

[Sid Belzberg]

Re: Here's Hoping They Don't Break Chess Next...

Agreed, not much different from how the human brain works, that was the point I was making. Now, the board games are where the engines shine because the board games all lend themselves to an eval() function in which 100% of the information that is needed is present and accounted for. The difference with Option Chess will be that an option saved for 20 or 30 moves from now cannot be evaluated (accurately), it can only be estimated much as stock prices and movements are estimated. It isn't about increasing the possibilities, it's about not knowing the possibilities accurately and oftentimes being WAY off. I hope you do realize this is a new concept that neither chess nor go encapsulate.

An even better example: the draft choices that football teams make from among a pool of college players. How many times do those choices turn out to be nothing like what was expected? More often than not is the truthful answer. A computer would do no better than humans at making those draft choices.

I'm not saying the engines won't play well, I'm saying they won't dominate humans the way current engines do in standard chess.

But yes, this has yet to be proven in the case of Option Chess. I'm quite confident that humans can and will match future computer engines.

i appreciate what you are saying but just as no human can take into account when someone will exercise these options a good chess player would likely continually take into account the two move possibility say 4 to five moves deep. In other words the human chess player is forced to think more tactically as strategic long term thinking is very unpredictable although i would guess certain patterns will emerge over time. Could an eval function be programmed that continually looks at the possibility of two moves consecutively deployed by the opponent or deploy two moves consecutively itself.?..yes of course. It would be a very different eval function then the current chess engines have which is why your variant is useful as a clever tool against online cheating with today's chess engines.
However I still maintain that the same technique used by Alpha go zero could be deployed in either standard chess or even option chess in programming only rules into a Alpha Go Zero like construct. The problem is that alpha Go zero had a world class opponent in its predecessor alpha Go to learn from but still was able to "catch up and surpass". So we would need someone really good at option chess for the thing to learn from.

[Vadim Tsypin]

Re: Here's Hoping They Don't Break Chess Next...

Chris, I took a few days to get and read Ender's Game (the original book) since I heard about it before but never actually held it in my hands.

Your comment makes me think: the siren song of AI appears strong. Driverless cars, applications like Hopper that tell me "Don't purchase this ticket now!" :), doctor-less diagnoses in fields like radiology and dermatology, now a robo-advisor that would manage my money better that I'd ever be capable to... - the world becomes so safe and secure for us, humans, doesn't it? Maybe The Singularity will result not in anti-utopia but in a genuine utopia where humans would be:

• relieved of mundane tasks and free to concentrate on creative pursuits ("our" prospective)
• happy and safe as toddlers playing in a secured sandbox (AI's prospective)

?

Horizons ETFs Launches Canada's First ETF Driven by A.I.

"A similar ETF launched last month in the U.S. The AI Powered Equity ETF (ticker AIEQ) uses IBM's Watson supercomputing technology to analyze data for approximately 6,000 U.S.-listed equities as well as regulatory filings, quarterly results, news articles and social media posts. The AI technology is able to distill this for the fund into 30 to 70 companies with the greatest potential for apprecciation."