Discerning chess plans with Stockfish?!

Intro

One of the most requested features of both Lichess and LiChess Tools is an automated chess teacher. And of course, people come with their own expectations about how teachers should behave... Real teachers ask questions, not provide the answers for you, though, so be careful what you ask for.

Anyway, the problem with any implementations you see out there, from amateur projects to chesscom, is that computers don't really form plans. They play chess probabilistically, so they can't explain the "best move" in a human language in a way that would make sense to a learning human. Most of the things you see there are other (often quite naive) types of analysis, turned into nice text, but yielding little learning value.

However, that doesn't mean computer engines can't be used to glimpse some ideas for the future.

Engine lines convergence

Some time ago I found a Lichess browser extension that did a very interesting thing: it highlighted the same moves in the engine evaluation. So of course I stole the idea and improved it in LiChess Tools. It's now high performance, the same moves are consistently highlighted the same way and you can choose to highlight just your moves or just the opponent's.

At first, I thought it was just a fun thing to do, but then I realized that it was providing very useful information. You see, if you configure the engine to display more lines (maximum of 5 on Lichess, but extensible to 10 with LiChess Tools for internal engines, BTW) and enable same move highlighting for yourself, you begin to see glimpses of converging lines.

For example, you see that all winning lines will eventually, maybe a lot of ply (half moves) down the line, move the queen to d4. That becomes the pivot of your analysis. "Why should I move the queen to d4? What does it do there? What risks does it involve and what can I do to minimize them? What can the opponent do against Qd4?". Immediately you start forming actual plans that start from an idea, a destination. The individual moves leading to Qd4 become less relevant and instead your thinking switches from "what moves to make?" to "how do I get there?".

Similar intuition can be extracted from common moves in losing lines. "Why does Qd4 lose when it seems so reasonable? What are the alternatives? Whatever I do, I must not move the queen do d4!".

Antithesis

Now, that's the optimal scenario. In many cases you get no common moves. In others the same move is shared between winning and losing lines, rendering it dubious, although you have to understand that any move other than the first in a computer line is considered best by the engine at that specific depth.

As I mentioned earlier, the engine does not make plans. The ability to intuit some ideas from line convergence is basically your skill, not the machine's. Even in situations where some plans are really well suited - like in known positions that have been analyzed and documented at length by chess professionals and theoreticians - the computer lines might not yield the same insights or none at all.

Sometimes the computer proposes lines that make little sense to a human mind, filled with positional complexities and exhaustive explorations instead of the easiest ways to convert the position. The blog post image shows this case, where Nf4+, followed by g6, is the common theme of most lines, but the first (and best by 2 centipawns :-/) is completely different. Even that is quite instructive: why is that move better? Is it? And if it's so complicated it cannot be understood, is it a good idea to use it?

Also, in some positions computer engines show the same line, with the moves in a slightly different order. This might trick you into thinking those moves are more natural or good plans, so there are caveats.

Example

You can see an example in a previous blog post of mine in which I was trying to analyze a single position. The arrows showing short to mid term plans were drawn using insights from the feature discussed above.

Let's examine the position more closely:

Only two moves are winning for Black. They are completely different, except the e5 push. Looking further down we see that other moves, even the disastrously bad ones, often are followed by e5. In this position e5 is hinting at a mid term plan. The non losing lines also contain a lot of immediate Nxf3+ and Bxh3. These are also the more direct winning ideas in the position.

I was saying in that blog post that the computer also wants to insert an h6 move that humans have not played, but look at the highlighting. h6 is a refinement in a single line that doesn't appear anywhere else. It may be the best move in the position, but it's not part of a plan.

Also, in my analysis, I've decided that the second best move as proposed by Stockfish is more likely to be the best for a human player. Look at the move highlighting: Bxf3 - first SF choice - appears only twice. Nxf3+ appears four times. If nothing else, one can decide if a move is more natural by counting the lines in which it appears.

Conclusion

While not revealing all chess secrets to you, the highlighting of the same moves in different lines generated by the computer is yet another tool in your analysis arsenal. The discourse changes ever so slightly from "I have to make these moves" to "The engine really wants this". When all else fails, this technique can provide you with plan insights in unclear positions.