Regan's conundrum

[jdart]

Thanks, that is somewhat helpful. The author did not clearly state the problem he was solving.

--Jon

[clumma]

Regan is replying to some of the points here in the comments on his blog, e.g.

https://rjlipton.wordpress.com/2016/12/ ... ment-79072

[clumma]

Are they derived from engine analysis of individual moves?

Yes

The moves humans do often have different long term goals from engine moves.

Maybe humans should adopt shorter-term goals if they want to play as well as engines...

Especially quiet moves.

In quiet positions many moves with approximately same eval. Human choosing any one of these will get small error score.

Maybe blunder analysis is more solid, but deriving human ratings on engine analysis of human moves in games is a bit far fetching, IMO.

How do you explain R2 fit of 0.996 to player Elo over millions of human games?

https://rjlipton.wordpress.com/2016/11/ ... g1612sc47/

Some advantages of IPR:

* Can analyze great players of the past (before Elo ratings)

* Proved hypothesis of "ratings inflation" false

* Can ask about playing strength over any length of time, including just a few games (still hundreds of datapoints)

* Can evaluate playing strength of a player in different phases of the game (opening, middlegame etc)

* Can be used to detect cheating

-Carl

[clumma]

Humans have certain goals. Some play for win, some for draw, some with contempt, some with respect. Strong humans up to these days have longer plans than even top engines

I suspect these plans assume nonoptimal play by opponent, or else engines would find them too. You can look at old games, like Karpov vs Korchnoi. The play seems so deep and logical, like a story. But it only seems that way when you miss key moves. Engine would defeat them like babies.

There was a time when "centaur" players could defeat engines. No longer. In almost all cases the human has nothing useful to add even when viewing engine analysis!

If Carlsen has contempt for an adversary, should we use Komodo +50 Contempt or Komodo 0 Contempt for analysis?

Contempt added to help engines score better in short tournaments against closely-matched rivals. Against humans > 400 Elo weaker, I would be amazed if it had any effect on IPRs.

Then the scaling of blunders. Also, never mentioned there is the compression of logistic with time control. And no mention of the phases of the game, which changes the shape of the logistic.

Fit of average eval error to Elo is linear. No logistic involved. Engine always analyzes at the same time control. It finds humans play worse at fast time controls, immediately before move 40, of course.

-Carl

[Lyudmil Tsvetkov]

the assumption that current top engines could possibly assess in any way top GM play is entirely wrong and inconsistent.

the only thing engines can see is obvious ( and shallow) tactical mistakes, nothing else. BUT, unfortunately, tactics is only 10% of what chess is all about, the remaining 90% are all positional stuff, involving quiet moves.

so, if you and Regan claim that this would be an analysis of 10% of the moves in the WC match, I am perfectly fine with it.

if you ask why top engines are winning top humans, very simple:

- humans get tired; if almost all computer-human matches, Kasparov-Junior, Kramnik-Fritz, even latest Komodo-human ones, the human either went into the lead in the first games, or was able to draw, and only late into the match the engine turned the tables. why? because the human got tired, while the engine is electiricity-fed to indulgence; this does not mean the engine played great moves, but only that the human started playing much lower than his abilities

- humans get distracted; well, evrything could possibly distract you while playing a game, some bad news, some camera shots, an unpleasant person appearing into the playing hall, or even this is just a bad day for you; computers do not have bad days, you know.

so, Carlsen actually plays 9 times better than SF, has 9 times better knowledge/understanding, but SF simply could win the match, because Carlsen will get tired at some point, Carlsen will make a mistake, etc. that does not mean that SF could evaluate Carlsen moves, but only that it will detect most obvious shallow tactical mistakes.

the picture is as follows: out of 30 moves played, Carlsen outplayed SF in 20 of them*3cps very small positional advantages=60cps, Carlsen has the advantage, but not winning, as SF does not make tactical mistakes; then, on move 30, Carlsen makes a decisive tactical mistake, worth 2 pawns, and loses the game. would it be rigth to assert SF played better? not of course, Carlsen chose the better moves in the vast majority of cases, outplaying SF, but he made a single losing mistake.

so I do not think it is a good approach to assess GM moves based on top engine output, tactical moves, yes, but assessing most quiet moves is simply not the thing to do.

[clumma]

Some advantages of IPR:
* Can analyze great players of the past (before Elo ratings)
* Proved hypothesis of "ratings inflation" false
* Can ask about playing strength over any length of time, including just a few games (still hundreds of datapoints)
* Can evaluate playing strength of a player in different phases of the game (opening, middlegame etc)
* Can be used to detect cheating

Oh, and it can be used to make engines play realistically at any Elo. To my knowledge nobody's implemented it yet. Stockfish's levels are sorta similar, but selects n-best moves randomly instead of keeping the running IPR to a fixed level.