Are there engine limitations that provide insight into importance of tactical training?

[chrisl64]

One of the perennial topics/arguments in human chess improvement is the importance of tactics and what proportion of training should focus on tactics (vs opening training and developing strategy/positional understanding). Obviously these things are intertwined and interrelated, but nonetheless I wonder if work in chess engines tells us anything or if the comparisons are just whilly inapt. I've seen various anecdotes about the strength of engines that have been limited in various way, some of which I don't really understand, and some of which may be true:

- Matthew Sadler discusses using "Leela at one node search depth" and, though he won 75% of the games, even with that limitation he lost 17 and drew 11 out of 101 games.

- Hikaru and many others have stated various limited engine accomplishments ranging from 1100-2300 FIDE. LazyBot is in the mid 2600s in rapid on Lichess. Many of these anecdotes might be the "telephone" effect at work, at best, if not outright fictions repeated as facts.

- I've found SimpleSearch and other projects using engines that are artificially limited, including in depth, but it's unclear to me what level of chess "knowledge" is built into the evaluation.

- But I also find discussions that essentially say "depth does not mean what you think it means" anyway, so the concept of an "x-ply" limitation doesn't make sense.

- Some folks use the strength of old chess computers as evidence about tactics, since there were no neural nets and the evaluation processes were somewhat rudimentary.

My question(s) are:

- To the extent that "tactics alone" -- which I take to be shorthand for something like "don't make simple one-move blunders, use simple tactics with minimal strategic consideration, don't miss one and two-move opportunities" -- means anything, are there projects that are relevant?

- Perhaps another way of putting the question: are there ways to limit engines in such a manner that their performance says something about the relative importance of tactics for human training (compared to opening training and positional/strategy learning, in particular)?

- If there is, has it been done? And what do the results mean/say?

(Sorry if title is misleading. I wasn't sure how to title this, as I'm not really a programmer)

[CRoberson]

Over a decade ago I invented an experiment to prove that all chess program strength/rating improvement was not due to speed and depth of search. After posting my results in this forum based on 1,000 games with Telepath, Bob Hyatt independently repeated the experiment with Crafty and 32,000 games.
It showed that taking out all the new eval features and backing the eval back to the 1980s drops at least 600 rating points off of engines that at the time were around 2600. The new derivative search 3 ply deeper, but was much weaker and only around 2000 Elo.

This suggests that tactics alone with ok opening knowledge will get you to 2000 Elo. There is a book "Rapid Chess Improvement: a study plan for adult players" by Michael de la Maza which shows the method he used to go from 1321 to 2041 USCF in 2 years using only tactics training.
He was in his 30s when he did this. I purchased the book and followed the trainining schedule for 6 months when I was 52. I went from 1370 to 1682 USCF in one year. I've heard many GM's state if you are below 1800, you need to work on tactics. Some put that rating at 2000 and
others 1900, but all I've heard are in the range of 1800 to 2000.

However, I don't agree with doing things like "puzzle rush": a timed suite of tactics. The book suggests a slower approach that gradually speeds up. Around 30 years ago, GM Lev Alburt published "Chess training pocket book: 300 most important positions and ideas". So, I am with quality over quantity on this.