Slashdot Mirror
Google's DeepMind AI Becomes a Superhuman Chess Player In a Few Hours (theverge.com)
An anonymous reader quotes a report from The Verge: In a new paper published this week, DeepMind describes how a descendant of the AI program that first conquered the board game Go has taught itself to play a number of other games at a superhuman level. After eight hours of self-play, the program bested the AI that first beat the human world Go champion; and after four hours of training, it beat the current world champion chess-playing program, Stockfish. Then for a victory lap, it trained for just two hours and polished off one of the world's best shogi-playing programs named Elmo (shogi being a Japanese version of chess that's played on a bigger board). One of the key advances here is that the new AI program, named AlphaZero, wasn't specifically designed to play any of these games. In each case, it was given some basic rules (like how knights move in chess, and so on) but was programmed with no other strategies or tactics. It simply got better by playing itself over and over again at an accelerated pace -- a method of training AI known as "reinforcement learning."
The only winning move, is not to play
So rise up, all ye lost ones, as one, we'll claw the clouds.
Please have it learn how to play modern strategy games like Starcraft and Civilization so we can have computer players which don't suck without massive bonuses which change the dynamic of the game.
-- All that is necessary for the triumph of evil is that good men do nothing. -- Edmund Burke
Reinforcement Learning systems have a tenancies of creating "Superstition" artifacts, were actions that may not create a net positive or negative are used over when the net outcome is positive. It often creates less than ideal outcome, but still it works. So this could mean a really long chess game with non-strategic moves, as the most optimal path, may not be enforced correctly.
If something is so important that you feel the need to post it on the internet... It probably isn't that important.
The world's gonna be an... interesting... place once someone merges this sort of code with virus code.
Check your premises.
No, not this one. Not even the next one. The one after that? Or after that?
Eventually, they will. The question is simply how long will that be. Right now, the ML pace continues to accelerate. Soon, they'll be stacking one skill upon another. The skill to walk. The skill to understand plumbing joints and leaks. The skill to know home construction. Etc.
It's coming. That whole "will never be able to" business... that's not going to pan out for anyone.
I've fallen off your lawn, and I can't get up.
Well, the program playing itself is not really qualitatively different than "if I do this, and he does that, and I do the other, and he does........ then I win!"; it's just carried out to more steps than a human would (because a human can't go that far). Therefore, any approach I can conceive of to go from knowing the rules to knowing how to win is pretty much equivalent to "running some iterations". Even the ability of human chess masters to perceive the board as a pattern instead of just a bunch of individual piece positions is probably approximated by something in the program.
Given that, I am unable to come up with a mechanism to go from "knows the rules" to "knows how to win a game" without doing something equivalent to "running iterations"...