Objectively Speaking, Is Magnus a Patzer Compared to StockFish and AlphaZero?

The reason I ask is I read this explanation of how it works:

http://tim.hibal.org/blog/alpha-zero-how-and-why-it-works/

and I don't see anything where it says anything about evaluating a position any way other than by "what worked in the past"....

What's even more impressive is that AlphaZero was just a way to test DeepMind's self-learning ability.

The Google team was essentially saying, "How about chess? Can it learn and master chess on its own? Hmm. Yes, it can. Okay, what's next?"

I don't expect to see any more games from AlphaZero, because it's already fulfilled its purpose (testing the AI). Now DeepMind will be probably be applied to other things. (Though, I hope I'm wrong on this, because I'd sure like to see more AlphaZero chess!)

Now look at that last sentence.  It is a sentiment that is probably echoed by Top Grandmasters all the way down to lowly patzers.  

I have read somewhere that Grandmasters were excited by AlphaZero's accomplishment because they could learn from AlphaZero!  This is the Silver Lining in the Eeyore Cloud!

Now didn't former World Champion Max Euwe write a book about, or titled, "Chess Master vs. Chess Amateur" where he annotated games between chess amateurs and chess masters, so that chess amateurs (patzers) could learn from those games?

With the same idea, why don't we have a match between World Champion Magnus Carlsen versus AlphaZero?  THEN, we could have someone write a book like Max Euwe, and then all us patzers could learn from World Champion Patzer Magnus Carlsen's blunders!!  Wouldn't that be a great learning method!!?

As long as the author of that book isn't Lyudmil Tsvetkov...

Especially you.

Well, Patzer Svidler is simply uninformed, as Alpha had its opening knowledge tuned on countless top human games, mainly the winning ones.

This whole story is not true:

- Alpha started from a more advanced code base, which is obvious

- and it did have a simulated opening book, further highlighting unequal conditions

But there is no better player who can comment Alpha-Carlsen games.

According to Google, the only knowledge given to AlphaZero was the legality of how each piece moves. No opening books. No games database. No evaluation algorithms or piece values.

Just "Here is how the pawn moves. Here is how the knight moves. The king can't move while in check..." (etc.)

AlphaZero figured out the rest by playing against itself and learning along the way. Hence why it's called "self-learning".

There may be some confusion here. The previous go program AlphaGo was trained using professional go players games. This program was bested by AlphaGo Zero which learnt entirely from scratch, and the same technique has now been used for chess and shogi.

We would not be right that often.

But yes, based on my understanding of the technology, it's positional evaluation network would be so good that without any explicit analysis at all it would play quite good chess. I am not sure how good it would be in this mode, but I do know it needs to do analysis to play at better than 2900 Elo (as it achieved near this level using about 1/30 of a second per move and got better as the time increased).

Yeah, then it rogered Stockfish.

how did this topic get posted just 1-2 minutes ago?

Lol.  ;-)

You can sell that to your mum.

They should have at least an algorithm in their cose base, instructing to rais values for pieces, when such are captured and the game is lost, and values for psqt, when piece placements on more advanced and central squares work well and win games.

This is a very basic autotuner, that can get to around 2800, and then will not improve further.

This is not a much advanced software, they don't have the code to distinguish between sophisticated evaluation patterns, and you can not advance much in chess without that.

Again, that was a colossal hardware: SF on it, if adapted, would be 3700.

So who is stronger now?

As per Talkchess poll, where all of the programming community dwells, believers and non-believers are split 50/50, so just the usual human pattern.

I'm not sure you grasp the concept of tabula rasa ("blank slate") learning.

The only input AlphaZero is given is the rules of the game. It then plays randomly, millions of times, against itself, learning from its mistakes and creating its own ideas and rules along the way.

It did this same thing with Go, and its learning progress was well documented. (In addition to chess, it's also the strongest Go player on earth now, too.)

Through this process, AlphaZero creates its own algorithms. This is why it bested Stockfish—Stockfish is only as strong as its human coding allows it to be. It calculates based on the piece values and principles written into it by humans.

AlphaZero, on the other hand, comes up with its own principles and values, figuring out its own ways to think and win. It's pure machine-learning, no human ideas or values involved.

Whatever AlphaZero is evaluating, it's obviously different from the way normal chess engines evaluate a position (using numerical pawn values), as AlphaZero seems to realize its winning long before the numerical evaluation on Stockfish does.

You can give Stockfish all the computing power you want, its playing will still be limited by the human-based principles and values written into it.

In several of the games, you can see Stockfish fighting over parts of the board that, due to its human programming, it believed it should fight for. Meanwhile, AlphaZero was doing its own thing, pushing Stockfish's pieces out of alignment, creating weaknesses, understanding long-term strategies that Stockfish, even with its superior calculating ability, couldn't grasp.

If anything, I'd say this shows that the future of chess engines will be based on self-learning AI, rather than human-programmed values.

I will stop arguing here, because it is meaningless.

In order to operate, each and every program should have its code base, do you understand that?

You want to tell me that, some Alpha just arrived from somewhere, installed itself on the TPU and started improving its play?

There is code guiding its actions, that is so obvious, code, written by humans.

Whether it fulfills its task on a single or multiple levels is fully irrelevant: it still does so following the instructions of the initial code base.

What do you think they are doing, when Alpha reaches an optimum and can not improve any more, they leave it to get things straight by self-learning?

Of course, they are changing the code base, trying to optimise it.

If it does not have instructions that winning is good, how could it evaluate then if a position is good or bad? Of course, it knows winning is good and that is WRITTEN in the primary code by a human.

You think it does not have instructions to learn where the pieces land? Of course it does. If it can not make distinction between different board squares, how can it then optinise its algorithms? So, it checks the squares where the pieces have landed during the game and, depending on the result, increases or decreases their values. This is still done according to the instruction that winning is good and that psqt should be increased in case of a win. That second instruction has also been written by a human.

So that, it is humans who wrote the primary code base and are constantly changing/optimising it, while the computer just follows those instructions. Even the indication that after each games colours should be reversed is written by a human. Is not that so obvious?

So, basically, Alpha just follows instructions, both during play and self-training.

World class GMs, professionals who use Stockfish every day, are absolutely in awe of the depth and beauty of Alpha Zero's games. The computer didn't just destroy Stockfish, it did it in style, rewriting some chess theory in the process! 

Some of those games were spectacular!