Fortress vs Stockfish 11

Wooow, very interesting data

According to that graph, for people who play stockfish in smartphone or bad computers, they will reach 3500 elo after reaching more or less 4 Million nodes. So, they will have to let their machines to think up to 4 Million nodes to produce that strength, it could be more than a minute. That's totally interesting.

You guys should take into account that I may be wrong in my opinion. I don't know so much about chess engines, but the little I know is what I speak. Usually I play them at default configuration, except maximizing number of threads and optimizing hash table.

I have chessbase, so I don't know about those features. I guess default is what you call infinite analysis. It just runs until you make a move or tell it to stop.

Sure, the engine's #1 move can change, even on the same hardware... but if stockfish doesn't like white's position on move 15, then it's reasonable to assume that it will disagree with at least one of white's moves.

Anyway, anyone can post a screenshot of what their setup prefers for move 17. And like I said, it never liked capturing with the king, so I wasn't trying to be tricky by only showing it when it liked capturing with the rook.

Since we have enormous databases of low rated player's games available (and by we, I mean people like Erik) it would be interesting to use machine learning to make an engine that really does play like a 1000, 1200, 1400, etc.

We are able to make really realistic fake human players... just need to find a market for it.

Do people with an interest in engines (like you) ever talk about this?

Programmers  already added random blunders to machine to simulate human (Add more blunders to lower rated elo , e.g. 3% blunders for 2500, 20% blunders for 1200)

For Stockfish, many people still feel it does not look like human as Stockfish's core evaluation/style doest change ( except more blunders in lower rating). In chessmaster series, they change both core evaluations and depth to simulate human with 150 different personalities. (. e.g,  overvalued bishop, overvalued queen etc)

There is machine learning Leela, in which she learn as human in earlier games, she blunder as human.

I played some of the old chessmaster engines, with the many personalities, and from what I remember they weren't very human... but it's been many years, maybe my memory of it is not so good.

I assume it's still artificial because ab engines are always based on calculation.

Humans do things like... you add pressure to d4, and even if it's unnecessary, just to feel safe, a human will play h3 to prevent Bg4 pin.

Or in other cases h3 is definitely necessary, but since it gives black an attack, a human will avoid playing it because it can't calculate a defense.

These are things you can't consistently simulate by limiting the search depth.

Humans find some 2 move tactics, but others are very hard to see.

Tactics that involve retreats, in between moves, quiet moves, things like this are harder for humans to see. Sometimes even when the position is on the board, and all the human needs is 1 ply search depth, they will be fooled into playing their original line.

So, for example, some simple rules to simulate a lower rated player:

80% chance of playing a check if there is a check that doesn't immediately lose material
10% chance of playing a check even if it immediately loses materiel.

80% chance of playing a capture if it doesn't immediately lose material.
5% chance of playing a capture even if it immediately loses material.

90% chance of recapturing regardless of the situation.

If an undefended piece or pawn can be won by making 2 moves in a row, 50% chance of making the first move.

If the last 4 ply were played on one side of the board, 80% chance next move will be played on that side of the board.

---

Obviously machine learning would nearly flawlessly make an enormous list of rules to accurately simulate players of all strengths.

There would be no reason for Eric to spend money on that.

Probably not good enough reason for Erik, but not "no" reason, and Erik isn't the only one who could do this.

You could market a computer that mimics various strengths and styles of play as a way for people to practice.

I mean... right now is a good example. Lets say people can't go to OTB tournaments. Well imagine you can play against a pretty good copy of someone who usually gives you trouble OTB. You set their rating, style, and opening preference, and you can play a nearly real tournament game against them.

And sure, right now the cost is probably too high.

But in the future... it would be pretty cool. I can imagine someone starting up a business for this.

Rubbish games- many of them. The whole TCEC concept is flawed- you cannot get top quality games under these playing conditions.

Take for example this one:

Only a total idiot would play 20...h4?? and suffer for the rest of the game- and yet, Stockfish played a move that no strong GM, or ICCF player would even think about: It is an obvious positional blunder of the first order.

Well then maybe your settings may be different from what I played. Anyways this was just a showcase of fortress concept and not an example of someone playing well against Stockfish. Anyone who believes it wasnt Stockfish is free to do so (that was not the main point of posting this game anyway), others may see the fortress example and gain something from it. This was just one amongst 100s of games I played against Stockfish (Android Droidfish) so it wasnt a claim of superior play.

@IMpfren, would you mean to say that top engines would walk into a fortress or bad positional play? If a top engine has a win or any means of avoiding a fortress in any position, imho it would go for the win. With all due respect, all games that you can point out where a top engines plays bad positional moves would be engine vs. engine games. Imho there would be no games where a top engine is forced into bad moves by a human no matter however strong.

@pfren  It is a known infamous critical positional blunder of Stockfish and it is not due to TCEC setting. Stockfish programmers are aware of that. Programmers can add fortress detection code to avoid such mistakes. However,  adding such codes slow down Stockfish performance speed and subsequently lose elo in general. 

 It will slow down approx 2% per code, and if you add 100 codes, it will be -200 elo  weaker Stockfish in bullet tests. It may equally perform in Long time correspondence chess though. e.g Crystal engine is a Stockfish fork with fortress detection code.

https://groups.google.com/forum/m/#!topic/fishcooking/UHbdD4DYc-Q

Look at their discussion presented by ICCF champion and SF programmer.

Don't worry, only one person said you were not playing vs Stockfish. That doesn't mean that the whole world agrees with him.

Only 5% of people will agree that Stockfish is not Stockfish

There are hundreds of Millions of people battling vs the best engines in the world (Stockfish, Komodo, etc.).

It's statistically impossible not to have from time to time some human players who won or draw vs these engines.

Hundreds of millions might play chess from time to time, but probably not even one in 10 of those will ever play against strong engine like Stockfish. Even that is most likely generous (maybe 1 in 50 is more likely).  The reason is simple, the majority of players are weaker casual players who play chess every once in in a long while, and when they do play, they will play against someone they know or perhaps online, certainly not against a strong engine. It's no fun getting crushed. Chess.com is a good example. I mean one in four players have a higher rapid rating than myself, and I have been playing for 2 months now, after hiatus of 20 or so years and my elementary school (and then I was just a casual player as well). 
Those who are most likely to get a draw or perhaps a win are titled players, and most of them will not be able to do so. The difference between Stockfish and a CM is probably bigger than difference between CM and some weak casual player.

Based on those facts, everyone is entitled to form their own opinion. That being said, in case you really did it (and the topic starter), well done.

It's true that weak players wouldn't enjoy by being constantly crushed by a powerful engine over and over again.

But on the other hand, I'm sure all of them tried at least once, under which motivation? just out of curiosity. To try is for free. So, that's where the hundreds of Millions of tries come from.

Not only that, one learns much more when losing vs superior level engines than winning vs weak players.

As an example, I watched in a documentary that Magnus Carlsen, when playing as a kid vs his own father, his father was always playing at full strength not ever trying to motivate him with a single win. Maybe that's the reason he got so good.

Anyways, original OP topic is about an example of forming and maintaining a fortress no matter if it's vs engine or human or whatever (anyways beating Stockfish became so common that it's not so impressive anymore). And this fortress is very impressive.

I'm used to see elementary endgame fortresses with small amount of pawns to save oneself from losing with material desadvantage, so this one big fortress has being really impressive and beautiful.

Most of those hundreds of millions casual players, would only play against their friends because they play for fun, not for improving that much. I have already said, on chess.com, I am in top 25% in rapid, and I am very weak and playing for 2 months after years of rustiness. . So if I am not that bad compared to average chess.com user (but bad in general), it is easy to conclude that majority of casual players plays few games a year not even looking to improve, just playing with their friends or family. Most of them do not even think about playing chess engines.

It is good to play against stronger human player if you wish to improve, it is true, but even strong super GM-s do not understand some engine moves, let alone weak amateurs who wishes to play against their friends or family and then leave chess alone for many months until another game. It is much better to have a friend who has, let's say 2 000 Elo points, because playing him will prepare you to face other strong opponents, as you will mostly be playing with humans not engines. 

Thanks mate, you summed it pretty well. That was precisely the intention behind posting this game. Fortress was primary, I added Stockfish only to provide credibility of it being against a strong opposition so that it's taken seriously. I am an above average player but can't claim to beat even an FM/CM, let alone Stockfish. This game just happened and somehow came out well for me. I created an account and posted so that the game is not lost and some people can see and enjoy it. That was the spirit. I won't get anywhere in life creating a fake post about drawing Stockfish if I haven't actually did so. Cheers.