This is a tricky one. I'm going to use a rather random example I had trouble with. YMMV and if you see it without any trouble, still try to understand what I want to convey.
We are searching for the rules that govern the acquisition of understanding in the most effective way. First of all, we must abandon all rules that are commonplace and that are just a waste of time.
- don't bother about visualisation
- don't waste your time with feeling the need to make an effort
- you are in learning modus here, so don't waste your time with trying to find the solution yourself, because you need elements that you haven't learned yet. There is no need to invent them yourself.
- when the tree of analysis becomes bushy, trial and error is a waste of time
What you need is to get an understanding of what you are trying to accomplish in a position. It starts with the goal. When you know the end goal, you can judge for every move whether it contributes to the goal or not.
 |
| White to move |
2r1r1k1/ppp1bppp/8/n4b2/2PPqB2/P4N2/4BPPP/R1Q1R1K1 w - - 0 18
What is the training plan? We start with the solution. If you see the solution, this problem is too easy for you. Let's assume you don't see the solution.
You don't see the solution, because you lack some patterns. So it makes no sense to press yourself to come up with a solution. There is no need to invent every wheel yourself. As a matter of fact, that is a waste of time.
Step 1: imagine what the end goal might look like. Don't get distracted by my formulation, let Stockfish help you.
The end goal :
 |
| White to move |
The end goal is a double attack on Na5 and Bf5.
Step 2. Imagine what is needed to obtain that goal:
- clear the e-file with tempo
- chase away the defender of f5 with tempo
Here you see how illogical it is to start with CCT. Of course you need tempi to execute your plan. But you need a plan first. Trying to find a plan while randomly looking at tempo moves is the other way around. Of course that might work, occasionally. Just as trial and error might work occasionally. But we are trying to optimize our training method. We are still in the study room. So we start at the beginning. Which is the end. If you know what I mean.
We must take away the squares on the diagonal b1-f5 from the black queen. Effectively we talk about c2 and d3. This leads to the first candidate move 1.Bd1. This is a multipurpose move.
- it clears the e-file
- it comes with tempo
- it takes away c2
1. ... Qd3
Step 3.
The black queen still defends Bf5. Can we chase her away?
2.Re3 is the only move that does that without any drawbacks, other than Qxc4 gaining a white pawn.
Step 4. Execute the plan.
After 2.Re3 Qxc4 you can execute the initial plan
When you have completed the four steps, you will notice:
- the solution is no longer complex. Since it is logical.
- since the solution is simple, it is easy to recall. Because you can always reconstruct the missing parts.
- now you can visualize the whole line with ease. It is silly to try it the other way around.
This logical narrative is what you want from the problem. But you can obtain a lot more.
Distraction.
You might get distracted by the fact that you might be able to trap the black queen. Take your time, and work out the logical steps how the queen will always be able to escape.
Transfer.
Now you have found the logical narrative, it is the right moment to ask: how can I get faster to the solution in future similar positions.
You start with LPDO. The knight on a5 is simple begging to be caught. How can you catch an LPDO-piece? There are two methods:
- trap it
- make it a part of a double attack
In this particular case, you are not going to trap the beast. But a second potential target readily reports itself: Bf5. From here on, you can start to build your narrative.
The training has now become a more verbal exercise. A linguistic exercise. You start with developing a verbal logical narrative. And then you play with it. Until you internalize the narrative. And when you become too lazy to verbalize the matters because you already know what it is all about by just looking at it, you are ready.
ReplyDeleteTraining is about feeding your inner LLM, so that it can generate thoughts of its own. That is my interpretation of prof Elan Barenholz' theory. It's like feeding your inner Lazybot.
Do you have information on the Lazybot pattern recognition training? I'm curious as to whether there was any connection between those training patterns and the typical human training patterns. If so, then there should be some way to create a teaching assistant that could provide training exercises and evaluations in terms of those patterns. I suspect that is NOT the case.
DeleteI haven't dived into that yet, but I had the same idea.
DeleteI'm thinking about a FEN-viewer with some extensions. Where you can show the LoAs, the PoPs, the targets et cetera at will.
This position is a good example of how the surface level features point to the essence of the position. Your arrows capture both the surface features and the essence. A double attack is required to take advantage of the specific configuration of pieces.
ReplyDeleteAs you noted, the first 'prong' is the LPDO BNa5. That can easily be seen by amateurs. The second 'prong' is the POTENTIALLY LPDO BBf5. It is currently 'protected' by BQe4 and NOT directly 'attacked' it; IE, the second component of the double attack resides at a deeper level. To reach that deeper level, we have to "SEE" that the "defender" (BQe4) has limited mobility to maneuver so as to maintain contact with f5.
Mentally remove the WBe2 and the BQe4 from the board. Now the goal of an ‘elementary’ Rook fork should become obvious. What is not quite as obvious is HOW to drive the BQe4 to a position that breaks contact with f5 AND does not create a new contact with a5. WHY? Because if either of the two ‘targets’ can be protected, then when ‘attacked’ the unprotected piece can be protected or moved away from the ‘attack’ in one move. Forcing moves (threats) are the mechanism to prevent this from happening.
There is a surface level “threat”: moving the WBe2 forces the BQe4 to move. Movement along the line e1-e8 does not counter the Rook’s attack. The BQe4 cannot safely capture along the 4th rank. By process of elimination, the BQe4 can only move to one of two squares along the b1-e4 diagonal. This restricted range of motion MAY trigger the hunt for a way to trap the Black queen. As you noted, this is a phantasm. Backing out of that illusion provokes the question as to which square the BQe4 can go to: d3 or c2.
That is decided by the movement of the WBe2. WBd3 is stupid, just throwing away the Bishop with nothing to show for it while freeing the BQe4. If Bf1, then the c2-square is ‘safe’ for BQc2. If that is followed by QxQc2 then the BBf5 recapture on c2 eliminates the potential double attack on a5-f5. If Bd1, then the d3-square is ‘safe’ temporarily. The followup is to ‘attack’ the BQd3 with the Rook. This maintains the initiative (preventing movement of the intended targets or protection of them). The BQd3 must capture on c4.
This requires “SEEing” that White is NOT forced to capture on c4 because his queen is protected by WRa1. The original double attack can now be carried out by WRe5. Black can attempt a Zwischenzug delay by capturing QxQd1+ but that does not change the dynamic between a5 and f5. After WRa1xd1, the double attack remains and Black will lose one of the pieces.
The point of all this verbiage is NOT to provide a “thought process” that must be absorbed into the subconscious. It is to show that there is considerably more going on “behind the scenes” than merely “SEEing” the surface level features (LPDO Na5 and potentially LPDO Bf5) and instantly grasping what must be done.
All of the considerations I’ve verbally noted must be somewhere in the internal LLM (using Barenholz’ theory) through training. If the LLM has never been trained on one of those critical elements, then it cannot “generate” the solution. Instead, it will flail around like a typical chess amateur.
Correction:
ReplyDeleteBlack can attempt a Zwischenzug delay by capturing QxQd1+
should read:
Black can attempt a Zwischenzug delay by capturing QxQc1+
Correction:
ReplyDeleteAfter WRa1xd1, the double attack remains and Black will lose one of the pieces.
should read:
After WRa1xc1, the double attack remains and Black will lose one of the pieces.
Since 1.Bd1, no Qc2 is possible anymore.
DeleteAfter a few weeks focusing on understanding, it will feel like learning a new language. Juan está en el jardín.
ReplyDeleteOnce the full pruning power of logic is discovered, you never want to go back to trial and error or CCT.
The linguistic science had entered a dead end with the ideas of Chomsky.
ReplyDeleteThe Status of Grammatical Rules
Chomsky: Rules are the cause and the fundamental source from which language is constructed.
Barenholtz: Rules are a consequence (epiphenomenon); an abstract pattern that becomes visible in statistical data after the fact.
Hence our quest for rules that govern chess must be doomed beforehand. You cannot create chess from rules. So I guess I owe John Watson an apology.
What is stored in our brain are relations between words. The words in themselves have no meaning. The meaning of a word is encrypted in its relations to other words. That's why grammar feels so artificial.
I have the feeling that I reach an endpoint here. I managed to fulfill the goal of this blog. All questions are answered which were raised along the way. I systematically removed the chaff from the corn. I can explain what works and what not and why. It turns out that the chess training world is an LLM that produces a lot of hallucinations. I have put the hallucinations to the test.
ReplyDeleteI finally have the feeling that I have optimized my training method. Not to its maximum potential, but to the limit of what we know today. I know what need to be done and how.
From here we enter terra incognita. Finally I am free to think for myself without the load of past theory. I'm curious what I will find!