Google plays Google learns

[Skybird]

https://arxiv.org/pdf/1712.01815.pdf

A new AI by the Google team has surpassed the great successes it had with its last Alpha Zero Go software, which taught itself Go by just having the rules explained to it and then playing Go against itself - after 24 hours it beat one of the best human Go players in the world. repeatedly.

The new algorithm optimizes the learning process and uses learning-optimised hardware components. This time it learned - by playing against itself - three games simultaneously, amongst my beloved poor chess. After four hours, the new Alpha Zero played chess so well that it beat the best chess software their currently is, stockfish, in a match over 100 games, by 28 victories and 78 draws - no loss. I know stockfish, everybody playing chess and being interested in computer chess knows Stockfish. Its the strongest chess software there is. Eh, was. Play is absolutely on human world class niveau.

The makers of this algorithm say their AI is so good because it no longer is limited by the limits of human knowledge. It operates beyond these.

I predict if Alpha Zero New would play against humanity'S best chess player, mankind would loose. Clearly.

[em2nought]

Wonder when they'll stuff that AI into Google Now? I tried to get Robin to call me Master and let me call her Jeannie, but she wasn't having it. Might have to try Google Now or Cortana instead.

[aanker]

Earlier this morning I watched an excellent analysis of a game between AlphaZero vs Stockfish 8 by ChessNetwork here:


They played 100 games, AlphaZero who taught itself chess in only four hours lost 0 and won 28 games with 72 draws. Google has only released 10 games. AlphaZero bested one of the 'best' (Stockfish 8).

Jerry aka ChessNetwork is an IM and I have been following him for years. If you have 18 minutes, check out the video.

[Skybird]

Wowh, what stuns me, what completely perplexes me is when the video says Stockfish calculates 70 million positions per seocnd, which does not impress me, but that Alpha Zero only calculates 80 thousand positions per second. That is a message in its own.

I recall times when peope, thought that advanc ein chess porgramming would not be that deicisve, but hardware (speed) imnprovmeent - this would be the drive behind better playinbg chess computers. Today ee have overwhelmingly playing chess oftware of unbelievably good start5egic and psoit9nal gameplay on hardware that is not even especiaqlly< fast, like a 6 year old cellphone. The prophets were wrong. Different chess software coding was the decisive ingredinet to poush computer chess. Not quantity of calculations (called brute force or Shannon A), but quality of preselections and according criterions ("knoweldge", if you want, or Shannon B), were the more important part.

Unfortunately, all this has done chess no good swervice, the game on professional level has fundamentally changed, and imo not for the better. Prepared knowledge and just calling it up has become most important. That is why Manus Carlson is so popular in chess: becaswue he doe snot like this conserved way of playing, too, but likes to fall back to the needs of actually calculating and playing during the running game. And that is where his strengths lie. Its a less efficient and mor eromantic approach, yes. But it fits the human dimension better, I think.

More and more players demand that the starting position at beginning shall be given up and shoukd be shuffled at random ever yagme, to bypass the dead, canned and conserved opening theory that computer analysis has revolutionised and somewhat "completed", and turned into something more oerfect, but also more lifeless. Think of that! This far it has gone already. Myself , I have fallen into the habit to switch off the openign library when playing against my computer programs.

[aanker]

I usually play over the board chess person to person the old way because I have found that there are too many cheaters on the internet chess sites using chess engines.

I agree with your observations on modern day chess methodology and prefer calculating my own moves too.

There are a few of us who meet at my place once a week or so to play and it is fun... also keeps me out of trouble - ha!

[Skybird]

I played correspondence chess back in the days. I was very good at it. Table chess I also did back then, but I then had a break from chess that was too long, and I never recovered from that interruption - roputione gone, too much knowedge forgotten. Even steel will rust if never been used for too long time.

Its one of the things in my life I really regret and am sad about. I do not play table chess anymore, its too frustrating and brings too many regrets, I lost too much skill. And correspondence chess - I could as well play against a computer. The rise of the machines ruined correspondence chess completely.

[aanker]

ChessNetwork analyzes one more beautiful game by AlphaZero - French Defense