Re:The neural network of the Stockfish chess engine
by tialaramex on Wednesday January 13, 2021 @07:43AM
(#25761291)
Right!
A solved game is categorically different from merely teaching a machine to be better at it than humans are presently.
Cepheus [0] is approximately a solution to Heads-Up (two player) Limit (ie the amount you can bet is specified in the game, you don't get to just bet arbitrary money) Texas Hold 'em poker.
In contrast Libratus is an AI that plays Heads-Up No Limit Hold 'em very well against humans.
You can examine the Cepheus strategy for yourself, if you could memorize it (it's too complicated) and were capable of making truly random decisions (Poker is a game of chance and so your strategy needs random action) you could reproduce it and be exactly as good at the game as Cepheus is. You can examine individual strategy elements and reason about them. For example if Cepheus gets an 8 and a 3 and you open with a bet, it will call your bet if they're the same suit, otherwise it will fold. On the other hand if those suited cards were an 8 and a Queen, it would raise a bit more than nine times out of ten.
You can't do anything about it (you might be thinking surely knowing that e.g. Cepheus folds 8-3 off here is valuable so I can benefit from that, er, no, "hiding" that by sometimes playing it loses more money than Cepheus gives up by folding it that's why this is a perfect strategy), if playing Cepheus, even with an incrementally "more" approximate solution you're not going to win a significant amount of money reliably in reasonable time, that's why it's approximately solved.
But Libratus isn't like that, it's playing some strategy that we know beat world class human players, but a further incrementally better AI might crush Libratus just as badly.
A solved game is categorically different from merely teaching a machine to be better at it than humans are presently.
Cepheus [0] is approximately a solution to Heads-Up (two player) Limit (ie the amount you can bet is specified in the game, you don't get to just bet arbitrary money) Texas Hold 'em poker.
In contrast Libratus is an AI that plays Heads-Up No Limit Hold 'em very well against humans.
You can examine the Cepheus strategy for yourself, if you could memorize it (it's too complicated) and were capable of making truly random decisions (Poker is a game of chance and so your strategy needs random action) you could reproduce it and be exactly as good at the game as Cepheus is. You can examine individual strategy elements and reason about them. For example if Cepheus gets an 8 and a 3 and you open with a bet, it will call your bet if they're the same suit, otherwise it will fold. On the other hand if those suited cards were an 8 and a Queen, it would raise a bit more than nine times out of ten.
You can't do anything about it (you might be thinking surely knowing that e.g. Cepheus folds 8-3 off here is valuable so I can benefit from that, er, no, "hiding" that by sometimes playing it loses more money than Cepheus gives up by folding it that's why this is a perfect strategy), if playing Cepheus, even with an incrementally "more" approximate solution you're not going to win a significant amount of money reliably in reasonable time, that's why it's approximately solved.
But Libratus isn't like that, it's playing some strategy that we know beat world class human players, but a further incrementally better AI might crush Libratus just as badly.
[0] http://poker.srv.ualberta.ca/preflop