Does "Elo inflation" reflect a real improvement in skill of grandmasters?

Of the Tabletop, and other, lesser varieties.

Moderators: SecondTalon, Moderators General, Prelates

User avatar
Eebster the Great
Posts: 2770
Joined: Mon Nov 10, 2008 12:58 am UTC

Does "Elo inflation" reflect a real improvement in skill of grandmasters?

Postby Eebster the Great » Thu Aug 24, 2017 4:24 am UTC

I have long known that the average Elo rating can increase or decrease over time, but that for top players, it has clearly been increasing. While the Elo rating system is very nice mathematically under certain assumptions, real results of players entering and leaving with different ratings seems to promote a substantial inflation of typical grandmaster ratings.

But today I was reading the Wikipedia article, and apparently some have suggested that Elo inflation represents a real improvement of skill at the top level. (This reminds me of the arguments surrounding the Flynn Effect.) The idea is not at all spurious; analysis of games by players from the 19th century like Paul Morphy makes it clear that at least at that time, serious inaccuracies and even blunders were commonplace at that time. I think few would disagree that there has been progress since then. If it is true that the population of rated chess players as a whole has not seen that kind of improvement, but that grandmasters have, then "Elo inflation" reflects a real change in the relative strength of the best players.

But there are other possibilities. For instance, it may simply be that modern communication has allowed a much larger population of players to get a rating, including players who in the past would not have been strong enough to even bother. Additionally, while there is little disagreement that chess has advanced since the 19th century, progress over the course of the 20th and 21st centuries is more controversial, and it is not at all clear that super GMs over the past, say, 30 years have truly gotten hundreds of points better compared to the median. What is certain is that computers evaluate modern moves as better on average than older ones, but this may simply reflect the fact that the best players today all use computers to evaluate their play. Maybe players today are just trained to play more like computers, even though that doesn't always mean playing better in the Platonic sense.

Personally, I feel like the truth may have some of both. It seems almost inevitable that top players have indeed made huge and consistent strides over the decades. But the average player may also be getting better, with the proliferation of chess programs and clubs. On the other hand, Elo ratings have inflated so rapidly in recent times that I think this may be overrepresenting any real gains.

What are your thoughts?

KnightExemplar
Posts: 5492
Joined: Sun Dec 26, 2010 1:58 pm UTC

Re: Does "Elo inflation" reflect a real improvement in skill of grandmasters?

Postby KnightExemplar » Thu Aug 24, 2017 5:32 pm UTC

Elo was never designed to be consistent across generations of players. It was always a relativistic measure that converges upon the following ideal:

Probability of win = 1 / (1 + 10^((RatingStronger - RatingWeaker) / 400))

That's it. Every player is assumed to have that distribution, and then the ratings are adjusted to try to make this hold true. There's no reason to believe that this distribution even exists, or that the "Rating" of players remains relatively steady over time. But Elo assumes that. As you can see, RatingStronger is compared against RatingWeaker, and are "floating". There's no baseline. So Elo scores as a whole will always jump around.

The Elo methodology was probably the best made methodology of the 1950s, back when it was invented. But today, statisticians are creating new algorithms such as "Microsoft TrueSkill" and "Glicko2" which try to improve upon Elo (converging faster for example). I don't think anybody really expects these rating systems to remain steady over a long period of time.

-----------------

In any case, modern chess players play hyper-optimized openings that have been computer-analyzed such that mistakes are highly improbable. We have databases of thousands of grandmasters playing "Sicilian Opening" alone, let alone the many other valid openings of the game. When people make mistakes, you can have a computer analyze your mistakes and tell you where to improve. Its far easier to get better at chess today than it was even just 15 years ago.

I'd say that maybe a "Grade A" player, a strong amateur, probably would win against the Grandmasters of the 1920s. Chess today has very easy computer-aided training accessible for free. And the community's understanding of the game is far, far deeper. Even without a computer to train you, a huge number of opponents are using computer-analyzed strategies against you, teaching you very useful patterns.
First Strike +1/+1 and Indestructible.

rmsgrey
Posts: 3109
Joined: Wed Nov 16, 2011 6:35 pm UTC

Re: Does "Elo inflation" reflect a real improvement in skill of grandmasters?

Postby rmsgrey » Fri Aug 25, 2017 5:53 pm UTC

Yeah, you'd predict that the more people get ELO ratings, the lower the skill level of median rated players will be in absolute terms, so the higher the ratings of the top players will be, independent of any actual change in skill either as individuals, or over the herd.

User avatar
Eebster the Great
Posts: 2770
Joined: Mon Nov 10, 2008 12:58 am UTC

Re: Does "Elo inflation" reflect a real improvement in skill of grandmasters?

Postby Eebster the Great » Fri Aug 25, 2017 9:08 pm UTC

That seems to give two conflicting opinions about how closely related the Elo inflation is to actual improvement.

KnightExemplar
Posts: 5492
Joined: Sun Dec 26, 2010 1:58 pm UTC

Re: Does "Elo inflation" reflect a real improvement in skill of grandmasters?

Postby KnightExemplar » Fri Aug 25, 2017 10:25 pm UTC

Eebster the Great wrote:That seems to give two conflicting opinions about how closely related the Elo inflation is to actual improvement.


My opinion is "not at all, Elo wasn't designed for that".

We do have the win/loss database of chess games extending back that far. If you were to seriously analyze the subject of comparing grandmaster skill levels across time, I'd think that "The Method of Paired Comparisons" is what you're looking for.

Indeed, Elo was derived from the Method of Paired Comparisons, specifically the Bradley Terry Model.
First Strike +1/+1 and Indestructible.


Return to “Gaming”

Who is online

Users browsing this forum: No registered users and 8 guests