Hacker News Reader

Comment by wslh | Hacker News Reader

The answer is: we don't know yet.

I'm not downvoting you but this is against every evidence based on ELO ratings and all the history of human-human, computer-computer and human-computer games: Carlsen is going to lose against my phone and probably every future human will.

Let's wait until the end of May but maybe we're already at that point with go as well.

(We don't know yet if I can win a running race against a Ferrari, still I'm pretty sure about the outcome.)

I guess for go it depends more if they bothered to train AlphaGo sufficiently before the match. If it has been training at the same level as before it is probably well beyond human reach.

We do know, with essentially 100% certainty. Carlson's elo rating is mid 2800s and the top engines are high 3300s. A 500 pt rating difference implies he has less than a 5% chance of winning a single game, let alone a match.

Plus, he can play against an engine any time he wants. If he could beat them in some magically surprising way, would it make any sense for him to be keeping this amazing skill a secret from the world? No.

> the top engines are high 3300s.

Sources?

Seriously?

https://www.google.com.au/search?q=chess+engine+elo+rating

And we don't know "with 100% certainty", but we can calculate a projected win/loss probability using those ELO ratings[1]

A 3300 rated chess engine vs Carlsen at 2850 gives a 93% win probability to the chess computer.

That's roughly the same probability of an International Master raned player (ELO 2300) beating Carlsen.

The top ranked engine is Stockfish, rated 3500+

[1] http://www.bobnewell.net/nucleus/bnewell.php?itemid=279

You haven't read the comments in your links, they say tuat this ELO was not calibrated with chess players which means it is not an ELO.

Arpad Elo's rating system has no single defined reference pool of players. The way it works is that you take a pool of players, establish the rating that unknown new players will receive (1500 in some cases) and then the system evolves from there. There are many disparate pools of players using the Elo system for their ratings, across many games.

Just because it was not calibrated with humans does not mean it is not an "Elo" system.

ELO scores can only be compared across a single pool of players, and those scores you linked to are from computer v computer tournaments only. The absolute scores aren't comparable because you can effectively adjust the absolute score by adjusting the parameters you start the system off with.

As far as Im aware no engines compete against humans regularly enough to get an ELO score (presumably because watching a computer thrash human players for a few hundred games isnt that interesting).

Sure, the rating compression thing is a valid criticism.

There's little doubt they are much stronger than humans though. There are enough online games that if this wasn't so it would show up.

Those ratings are uncomparable, since humans only play humans and computers only play computers. It's probably way lower than 5%.

Carlsen himself certainly realizes he would have no chance in a match against the top computer engine. Its effective rating is probably at least 400 Elo above his.

We definitely know. Engines are taking handicaps when playing humans in modern tournaments.

https://en.m.wikipedia.org/wiki/Human–computer_chess_matches

Reply on news.ycombinator.com