Play with computer is too easy

@Sybotes said in #8:

But why, @Toadofsky ?

When I just go in analysis mode and let me show the best move, then engine shows good moves. For example, recapture a rook or something (even if I can win even then).
But in "practice" mode, the engine often makes moves that are not even among the best three Stockfish moves. And this only happens in the endgame.

It looks like the "practice" mode uses another algorithm to pick its moves in the endgame. Hence my suspicion it might be a tablebase move. And I really see no reason for that.

Fair question; I'm answering about what I know (the engine outside a browser), but if you want me to speculate...

Maybe analysis uses cloud evaluations and practice doesn't? I have no idea. https://lichess.org/blog/WN-gLzAAAKlI89Xn/thousands-of-stockfish-analysers

Thing is you don't need a cloud of stockfishes to see that Kxf4 is best in that puzzle. The other two moves staying near the pawn also make sense, but Kd2 makes no sense. Also, i can run the browser stockfish on the position in the study i made and it already comes up with the three moves I mentioned. So its quite mysterious why "play with computer" doesn't manage to find a reasonable move.

Here is yet another example of the weak (poor) moves made by "play with computer" starting from a completed endgame puzzle.

The puzzle is here:

https://lichess.org/training/9Yniu

After completing the puzzle the position is ofcourse completely winning for white, but say in a blitz game you still want to put up resistance because white's time is down. Black should make best moves to slow white down. However, if you click "play with computer" when the puzzle is done, the computer plays g5. Obviously the best chance here for black to prolong the game is to try and promote a pawn with a3. I made a study of the position and a3 is +5.4 for white. The move made by "play with computer" (g5) is +71! Its not even among the best 3 moves: a3 (+5.4), a5 (+61.9), and Kg6 (+70.9).

Try it yourself if you doubt me.

Here is the puzzle:

https://lichess.org/training/9Yniu

Here is the study:

@EmaciatedSpaniard said in #13:

Try it yourself if you doubt me.
Nobody doubts it. I gave you a link with an older discussion on this topic with many examples. But nobody explains it either.
@Toadofsky will just tell you you shouldn't use it on the Lichess website but on some console offline ... Well, there are wise people on the world who shouldn't be bothered by such noobs as you and me.

Thankyou for mentioning the previous thread. It seems one didn't get anywhere then.

Toadodsky maintains that the developers of SF don't care about this. But what about lichess? The button is on the lichess website in every puzzle. It is named "practice with computer" but there is no practice possible when the computer makes such weak moves. On the contrary, practice would probably be detrimental (at least for endgames). I'm sure someone cares about it - the people who were involved in designing the button to begin with.

@EmaciatedSpaniard said in #15:

Thankyou for mentioning the previous thread. It seems one didn't get anywhere then.

Toadofsky maintains that the developers of SF don't care about this. But what about lichess? The button is on the lichess website in every puzzle. It is named "practice with computer" but there is no practice possible when the computer makes such weak moves. On the contrary, practice would probably be detrimental (at least for endgames). I'm sure someone cares about it - the people who were involved in designing the button to begin with.

Indeed, that's the critical question, sadly one I'm not in the best position to answer (I love requesting features, but am only insistent/demanding with the team on concerns I think are of utmost concern, yet this concern seems potentially hairy).

I never insist or demand, but i think its ok to point out shortcomings of lichess features on the forum devoted to lichess feedback. I try to make my postings constructive and specific. I trust that the lichess developers can prioritize their time and ofcourse i want them to work on the most important/urgent issues first.

In this case, I wasn't aware there was a previous posting and my first try was ignored by the forum readership except for one joke comment. In this particular thread I was initially greeted with laughter and comments implying it was all 'in my mind'. So I persisted. Now I think together with your previous thread we have some solid examples illustrating the problem. Anyone who cares to can test it for themselves. Perhaps eventually it will get some attention by lichess.

In the meantime I just won't do any 'practice with computer' on endgame puzzles.

I took a look at that study, and the issue there is that stockfish in "play against the computer" mode has access to the 7 man endgame tables, and they are distance to conversion. Since there are winning pawn moves(which count as a conversion), the distance to conversion for any losing move will be 1 turn, so it sees all of the losing moves as equal. This issue has been discussed on the endgame database forums quite a bit, although mostly from the other side where the computer is winning and will try to get to a conversion as fast as possible even if that delays a win greatly. The classic case is a KQR vs K endgame where it will move the queen next to the enemy king to force it to take the queen because that is a conversion into a winning endgame. For human play, using distance to mate tables instead would be much better, even though their evaluation is sometimes wrong due to the 50 move rule. Or, when the tables say that it is losing, turn them off and try to defend using only search, which is unlikely to do worse :P

Very interesting, so there is basically no way to rank the different moves as they are all the same 'distance to conversion'. But in a chess game with a move count (or timer) what matters is distance to mate isn't it? I mean the losing side should try to prolong that move number and the winning side should try to shorten it.

I just started a study of Capablanca's classic endgame lessons (KP vs K, drawn) and discovered that "practice with computer" is even worse than i thought. Even when white still has a chance to promote a pawn if black makes an error, white throws the pawn in front of the black king and allows capture. This moves makes no sense even when consulting a tablebase does it? I mean white should make moves that preserve winning chances, shouldn't it?

Please see the 2nd chapter of this study "Lesson 1: KP vs K, drawn: Practice with computer" to test what I describe: