Should we analyse at much lower depth?

As I understand it, an engine looks at how good a move is, based on a number of moves down the line.  And each move in that line is the best move given the total depth.

So the best move at depth 20, is a move where, if afterwards, the opponent plays the best move, and then you play the best move at a depth of 19, and then 18 and so on (so the last moves have very low depth to their analysis), you will get the best outcome.

However it occurs to me, that I will never find/play those 20 moves, and if I look at the line, they can be non-human moves.

What I _really_ want, I think, is something that puts me in the best position in 4 moves. That I might realistically find.

Now, I am not talking about say an opening, where a certain move is main-line because it's been well tested and tried etc.  If you're going to memorise a best move from a position, then you need to know what comes afterwards to make it the best move.  Or if you want to analyse an error etc.  Especially middle/end game.

So my hypothesis would be that you should use an engine set to a depth you are capable of calculating in a real game.

As a relative newcomer to chess, I'm looking forward to hearing why that's wrong.

I always see/hear comments about "Oh, that was at a low depth", or "I set this to max depth"... 

As a programmer who studied maths to high depth (long ago), my hypothesis feels correct, but if that were the case, would it not be common practice for players < 2000 to lower the depth to something much lower than the default.

Please enlighten me

I'll ask the pixies at the end of the garden, they will know.

I was worried this might be a heated and fractious debate - but that's drowned out by the sound of the tumbleweed....

Although the specific idea isn't so good, your intuition is correct. For example professional players aren't as interested in the "best" move as they are the most practical move. They also take it a step further, for example with Caruana talking about looking for lines that involve him playing objectively poor moves to reach a position where his opponent is ahead, but requires inhuman play to prove it. That way he simultaneously sidestepped his opponent's preparation (they're less likely to have prepared against "poor" moves).

The reason I say your specific idea is wrong is largely for the same reason memorizing high-depth moves is wrong, and that is humans and engines play fundamentally different ways. In other words having it show you suggestions at low depth doesn't suddenly cause its suggestions to be any more practical or "human."

At the end of the day, the engine is best suited for analyzing positions that allow it to show its strength, which is calculation. In other words tactics. In strategic situations, at low depth especially, the engine sometimes suggests objectively wrong or impractical moves. Letting it think longer alleviates that, which is why people may criticize analysis done at a low depth. And yes, at high depth, the moves it shows you (especially if you're newer to chess) may be completely impractical even if they're theoretically best.

Many times after a game I'll be checking with the engine, and I feel bad for lower rated players... because my winning plan made it so easy, while the engine line is ridiculous. If they spent an hour with the engine they'd only get confused (or learn the wrong ideas) while a human could make it simple after only a few minutes.

(This is also why it's sometimes easy to tell when a person is cheating, when their method is impractically precise)

Anyway, this post is to say that it's much better to get a well respected book, and play through every line of analysis on a board while keeping notes in a notebook. That's what will teach you about chess. The engine is more for immediate surface level questions like "why does this or that move immediately fail."

A debate might get sparked if some people disagree... maybe try to get @tygxc in here... I'd probably post a few more times to argue about things if that happened

Unfortunately though there are a lot of kids these days, so utterly vacuous posts like #2 above are not unexpected.

I understand what you're saying.  However, is that really right?

For example, if you do a depth of 30, you might get a series of moves where a specific move is preparing for something much later.  Whereas you _remove_ that element at low depth.

To go the other extreme, if something is going to put me in the best position after 1 move, it's unlikely to be non-human, as the result of the move will be a better position in terms of the factors (I assume king safety/activity/material), and it's not looking to see "is this still safe in 10 moves time".

I do see that sometimes Levy will say something like "look how brutal the computer is... it wants..." and then show a 3 move combo that is non-human, but doesn't looking ahead fewer moves mean it is _less likely_ to be non-human, more likely to be something I can actually use/remember/arrive at.

I did do a little experiment with the position which threatens fried liver but where traxler can be played.

The Traxler requires knowledge of how to play it for at least 8 moves from here, across many variations.  d5 is the simple and most common response here.

At low levels, the engine was suggesting d5, and Bc5 was not in the top 5 suggestions.

As the depth increases, Traxler Bc5 moves up the list.  It never makes it to number 1, but again you need your opponent to know all the moves.

So this would support my theory, because if I was reviewing a game and wanted to know where I went wrong, what's the best move here, at lower level, I would expect to see d5.  Doing a maximum depth and seeing Bc5 and playing that would get me nowhere, if I don't know the Traxler lines....

Does that make sense?

I know a sample size of 1 is fairly meaningless, however it does seem to make sense to me at least

That's a good test, I like that. The important thing is the depth has to be low enough that it changes which moves it likes best before going outside of your preferred range. In other words I don't think we're talking depth 20 vs depth 30, more like depth sub-10 vs depth 10.

IIRC (and as @drmrboss likes to mention) the neural net engines (like leela) play somewhere around 2000 level when moving "instantly." I'm not sure how it can defend against tactics if the depth is zero, but sure, like you're saying, if it's valuing strategic elements like piece activity, king safely, space, and pawn structure, then its suggestion may be a useful guide for what types of moves you should have as candidates... I've never tested that so I can't say for sure.

I was writing before your second and third posts.

I agree with what you say for using engine for learning.

I tend to use online resources, eg videos and the like, and quite often there is only one good move in a position, or something that happens, where I just quickly use an engine line to see what I did wrong.

In those cases, I was wondering if low depth was better.  People complain that the game analysis doesn't show the "right moves" and the response is "it's low depth", but those best moves are more realistic, or "practical", than a deep engine one.  It doesn't mean they are all the time, of course not, however it's better.  So... if you're going to use the engine for something, set it to a depth based on your level, higher is not automatically better!

(theory, non-proven)

Exactly, whereas I see people saying "I ran an analysis at max-depth..." (seems to be about 40 online) and therefore they know the "best" and right answer.  If they play their next 20 moves perfectly.

If there's truly only one good move... well let's pretend it's also a forcing sequence (so there aren't many different lines to understand). In that case I think studying the true best line is the most instructive.

It's positions where there are many reasonable moves that are open to interpretation so to speak. Those are the ones you'll want to get creative with your engine or non-engine use. IMO.

When I want a high depth I fire up stockfish on my PC (stockfish and interfaces are free to download, just making sure you know).

But you may not realize why I want a high depth... it's that humans can take shortcuts in our thinking while engines have to get there move by move. By letting it reach a high depth, I'm  trying to guard against the engine's idea being strategically superficial.

Here's a fun example

-

-

In this Ruy Exchange structure, as a king and pawn endgame, white is winning.

White wins because white's pawn majority on the kingside can create a passed pawn but black's majority on the queenside can't. The simple defensive idea for white's queenside pawns is don't initiate a pawn trade. If black is always the first to play pawn takes pawn on the queenside, then black will never be able to make a passed pawn.

So the winning plan is simple:
1) Make a passed pawn
2) Push it down the board so that black's king has to chase it
3) With the black king distracted, you infiltrate with your king to win pawns

(If you practice this position, the winning plan is simple, but the execution vs an engine is not)

As humans we can understand this without making any moves... and what's more, is that this structure arises on move 4. In other words humans can "see" 50-100 moves ahead instantly in the sense that we know as the game goes on who is doing better, who needs to be avoiding which trades due to which endgames, etc. Engines can't know these things until they see them.

So depth 30 isn't so much about knowing what the engine had in mind for move 29. It's more waiting for the engine to catch up to your knowledge... of course the point is it's superior in the short term, so you're hoping to marry your long term ideas with its short term accuracy. This is how I (and I believe many other experienced players) use / see the engine's input.

Great post, thanks!