Tactics cannot be very important for chess

[jp]

Leela has zero "planning". Whatever the type, engines just calculate.

[Ovyron]

I think Stockfish's search is good enough to be able to become the Unbeatable Opponent people talk about, it just needs a strategic eval, and once it has it it will not need very fast hardware, and even the search could be stripped down and it'd still draw.

Just look at a game where Stockfish lost, the blunder was caused by a distant position that it wrongly evaluated, and all the search was wasted because it tried to refute this wrong eval from both sides and failed, so it became the mainline, the nonsensical mainline that lost it the game, because of a single position evaluated wrongly.

I think the only way to reach that goal would be to abandon the current elo chasing paradigm and focus on having a strategical evaluation. Once again, I'll suggest a "David Vs. Goliath" experiment, where Stockfish Depth 24 tries to hold his own against Depth 34, the blunder that loses it the game (when compared to a Depth 34 vs. Depth 34 game ending in a draw) can be examined and the eval improved so that lower depth is enough to hold the game.

But instead of going and checking how Stockfish loses a single game, people make a small tweak and play hundreds of thousands of games to see if it improves elo or not, without understanding why, because elo chasing is the norm, and an engine that doesn't lose any game ever but isn't good at winning either wouldn't have an outstanding elo, so they could even have it on their face and discard it.

[jp]

Just look at a game where Stockfish lost, the blunder was caused by a distant position that it wrongly evaluated, and all the search was wasted because it tried to refute this wrong eval from both sides and failed, so it became the mainline, the nonsensical mainline that lost it the game, because of a single position evaluated wrongly.

Can you give some examples of such games? Misevaluated at what depths? Obviously, all engines still lose games, so they must be making many mistakes, but you're making the cause of those mistakes sound very simple ("a single position evaluated wrongly").

But instead of going and checking how Stockfish loses a single game, people make a small tweak and play hundreds of thousands of games to see if it improves elo or not, without understanding why, because elo chasing is the norm, and an engine that doesn't lose any game ever but isn't good at winning either wouldn't have an outstanding elo, so they could even have it on their face and discard it.

They do that because if they just try to "fix" it based on one supposed blunder in one game, they'll weaken it in other ways and it'll show in other games. And how are they supposed to do this checking without using human judgement, which cannot be trusted?

[Ovyron]

Can you give some examples of such games? Misevaluated at what depths? Obviously, all engines still lose games, so they must be making many mistakes, but you're making the cause of those mistakes sound very simple ("a single position evaluated wrongly").

Huh, can you show one example where this does NOT happen?

Here's what basically happens when Stockfish plays a move:

-it sees a position
-it checks how much time it has on the clock
-it sets a timer that tells it how long to take for this move
-it builds a Principal Variation that is its best guess on how the game will continue if both players don't make mistakes
-it gets stuck on a loop where it keeps extending the variations move by move (called half-move, or Ply)
-it evaluates the very last move of the PV with some score (or with bounds that it knows the move goes over, or under)
-it goes to the tail of that variation and compares what it is with different moves from the penultimate side to move, then with 3 plies before the last move, then 4 plies before the last move, and so on, until reaching the root position again
-if it finds an improvement for either side it builds a line and scores it (or gives it some bounds)
-it stops when it can't find an improvement
-It checks if the timer ran out, if not, it estimates if going one Depth higher would go over the time, and if it would, it plays the move, otherwise, it goes back into the loop and extends its variations one depth higher
-If the timer ran out, it estimates if the position requires more time, and it might play a move or go back to the loop

That's it.

So if Stockfish played a game losing blunder it'll be because the evaluation at the tail of its PV was wrong. If it was correct a move that didn't lose the game would have been played.

They do that because if they just try to "fix" it based on one supposed blunder in one game

"Supposed blunders" don't exist. If a game was lost it's because a blunder happened. If no blunder happened the game ended in a draw.

they'll weaken it in other ways

How can a more accurate eval of positions weaken the engine?

And how are they supposed to do this checking without using human judgement, which cannot be trusted?

Because you have all the time in the world, you could check what moves were not game losing blunders, see what was the eval for them and why was it scored worse than the PV (there's a problem somewhere because a superior move was compared to the PV move and the engine wrongly concluded it was a worse variation.) You make the eval more accurate so the engine doesn't pick the wrong line when comparing moves, and it should not have any drawbacks.

[jp]

They do that because if they just try to "fix" it based on one supposed blunder in one game

"Supposed blunders" don't exist. If a game was lost it's because a blunder happened. If no blunder happened the game ended in a draw.

No, that's not a "blunder". That's a "mistake".

[jp]

they'll weaken it in other ways

How can a more accurate eval of positions weaken the engine?

You won't produce a "more accurate eval of positions". You'll at best produce a more accurate eval of that one leaf position you're blaming for the engine losing the game. You can of course make the evaluation function bigger and bigger, but that comes at a cost.

[mclane]

Leela has zero "planning". Whatever the type, engines just calculate.

But it plays more strategic then stockfish, Komodo etc.

[jp]

Well, there has not been agreement in this or previous threads on a sensible definition of "strategic" for engines.

[Ovyron]

Well, there has not been agreement in this or previous threads on a sensible definition of "strategic" for engines.

Okay, but the question is if we can come to an agreement? Why not?

I have said that if static eval finds a move without any calculation it's an strategic one, in most positions this will not be possible at all (at least, there could be many strategic moves that are equal, and a tactical one that is better than them, and you can't expect anything to find the tactical shot without any search), so it only makes sense to talk about strategic moves in quiet positions where there's nothing to do.

If with enough search all engines agree that a move is best, it's a tactical one.

Otherwise, if no engine can agree on a move, the one that would have a better performance against the rest is the strategic one, the others would have a bad strategic sense behind them, to the point of shuffling.

I have the worst track record when it comes to people agreeing with me, though, so you probably want someone else to propose something better for what a strategic move is.

[swami]

A strong GM, many years ago, wrote something to the effect "Chess is purely about tactics..."

His thinking was the positional play is nothing more than very deep / very long-range tactics. IE what is bad about a backward pawn? It gives you a weakness and your opponent a target. A target that can't move, which will be around for a long time, most likely being lost at some point. In that light, I have to agree with him. At some point in the future we won't need a backward pawn evaluation term because the search will see deeply enough to realize the pawn will be won or lost, period. I have seen a few evaluation terms in my programs over the years that became superfluous.

The ultimate goal is to checkmate the opponent's king. That is purely a tactical goal.

There is the notion that strategy is just deep tactics.
And as the searches get deeper, the line between tactics and strategy may blur.
But there are some moves that we make which really do not have a tactical goal at the time we make them.
But I guess the bottom line is to always choose the best move(s). Whatever that may be.

There is one aspect which is difficult to capture, and the one which is critical, in that it makes or breaks the game.

It is "Planning"
We are given with abundance of choices in an otherwise normal position, it is the "plan" that we decide to carry out, based on our intuition and decision making, which ultimately determines the outcome. This aspect is crucial to get the edge over competition. This is difficult for human to explain, and I suspect it is difficult to program, and certainly difficult for somebody to teach. This differs with every position in every game and it is hard to replicate and it may work in one game, and it wouldn't work in another, and this is what makes Chess a very challenging game, even in modern times.