Computers are now better than people at Chess, for example, but what are some games people are still better (at)?
The traditional answer is Go.. That article is a little old but I’m not aware of any breakthroughs since then.
Twister.
Civilization IV. Even at Deity level, where the computer’s given totally unfair advantages, a highly skilled player can defeat the computer nearly all the time.
And I specify CivIV because the A.I. in Civilization 5 is a total fucking joke. :rolleyes:
Mmm…let’s assume robotics are included.
Dungeons and Dragons.
This one is pretty interesting. Computers generally play turn based games with graph or tree searches*. This means the efficiency is largely based on how many “successor states” exist to any possible state, meaning: given the current way the board looks, after this turn, how many different ways can the board look? Chess starts out with, if I’m counting right, 20 successor states (move any pawn forward either one or two spaces = 16 + move knights two different ways each = 4 = 20). This number raises and lowers depending on the board, but generally it’s pretty good. Of course, there are a ton of states still, so we introduce little rules to nudge it into trying certain paths of computation instead of others.
Go’s successor states are… obnoxious. Go is played by dropping a single stone on an unoccupied square on a 19x19 board. The very naive calculation gives 3^361 states (each position can be black, white, or empty). That’s total states, but given how the game is played, one can see even the number of immediate successors is immense.
Really good tuning of which states to analyze isn’t just helpful, as in Chess, but necessary to even make the problem of beating a two year-old tractable.
This is, of course, assuming we don’t use a different method to make a good Go playing computer such as machine learning. I’m speaking purely of the traditional approaches we use for other games.
*I’m lumping things like minimax in “tree searches”
Rugby?
Sigh…
Just imagine we took an able-bodied human and replaced his brain with a computer.
Nah. All we’d have to do is invoke via Bluetooth the James Tiberius Kirk manoeuvre and tell him/it to compute pi to the last decimal place.
What does that mean, though? We don’t have any software that can run a human body, let alone manage and make sense of the sensory inputs. It’s only pretty recently that we’ve had bipedal robots that can run, or process a video feed and navigate a real environment.
Pretty much everything required for rugby is still a hard problem for computers.
Texas Hold 'em, considering the incomplete information.
I would be absolutely amazed if the best computers couldn’t consistently beat the best humans at Texas Hold 'Em. Yes, the information is incomplete, but it’s incomplete for the humans, too. The computer, meanwhile, can calculate all the probabilities for the missing information instantly.
In the long run you would be correct, but if the thread is about a human beating a computer in a game of Texas hold-em in a single series of games, then my money is on the human.
The computer knows all the probabilities, and thus will play very predictably. The human will not.
Chronos- No sooner than I typed than I realized you’re obviously correct.
Don’t think this is correct. Just like computers can have a huge database of previous chess games played by the opponent, they can have strategies used by opponents in Hold-em. The program doesn’t have to be entirely based on probability.
One interesting case is the Social Ultimatum Game, a modification of the Ultimatum Game. The Ultimatum Game is a boring game theory game, there are two steps: player 1 is given $10 and must make an offer to player 2 of some portion of the money. If player 2 accepts, they get the proposed amount and you get what remains, so if you offer $3, and they accept, you get $7 and they get $3. If they refuse the offer, nobody gets anything.
The optimal strategy as player 1 is to propose $1 and the optimal strategy as player 2 is to accept any value other than $0 because it’s better than nothing. Usually people will offer far more fair trades, or irrationally reject poor trades even though it ends the game. Computers are very good at acting rationally on this. What they suck at is an extended version of the game with multiple rounds and more than 2 people, because it allows people to hold grudges and be nice to each other.
That paper is about making an agent who can, essentially, learn to form alliances and strategically snub people. It actually does fairly well after an extended period of learning, but it’s still not quite as good as people are.
I think this characterizes a lot of the troubles computers have with poker. They need to have an internal model of bluffing and models of how good/bad others are at bluffing. If you could get a computer to model that on top of their ability to quickly compute probabilities, you could get an exceedingly good one. It’s not a trivial problem, but it will probably be solved eventually.
How about Bridge?