It’s going to be interesting to see how this affects online poker.
A lot of bots already exist, but bots that are consistently better than humans are really going to increase the arms race between poker sites and bot authors.
With Go and No-Limit Poker falling to bots I don’t think there are any games left where humans dominate bots, just games that bots haven’t been written for yet.
Ha! I can’t wait to see the Starcraft results. The complexity is way higher than these toy games that so simplify the interactions between combatants. I want to see it combine 3D object recognition, planning, replanning, detecting/making bluffs, and so on. These games, especially simple mechanics + billions of moves analyzed, make it too easy for the AI’s to outdo humans. I want to see harder situations closer to the messiness that is the real world. Then, see them do the same with as little data as we work with on the games themselves.
They’re writing the bots now.
I know. So far, though, all the best bots for Starcraft were defeated by humans trivially despite defeating other bots with amazing displays of micro or planning prowess. The humans just spotted their patterns then worked around them. Occasionally, they bluffed them to hilarious end.
An example was the bot who did battles between units based on a scoring mechanism to determine the strength of one, a group, and so on. Human sensed this. Counter was to send large group of units at an enemy group to make them scatter. Then, split that group into smaller ones to go after each individual enemy unit. Scoring made enemy units weaker than what was sent after them. So, enemy units always would flee without firing a single shot back at human players despite actually being able to defeat a bunch of them. Another example was even simpler where human just ran some units randomly through the enemy bases while building an army to hit them. Enemy AI was unprepared as it wasted all its focus countering the “attack” since it couldn’t call a bluff.
It’s not StarCraft-specific, but here’s some progress on the graphics recognition you wanted: https://arxiv.org/abs/1805.11592
That’s really neat. Thanks! On the other end of things, there’s some circumstantial evidence that my prediction in favor of humans on Starcraft will be right.
I agree with Nick that we’re far from bots dominating all games. I’m not sure Starcraft will hold out that long once they get going on it because computers have an inherent edge in the reflex-based/micro parts of the game.
I’d love to see someone work on competitive bots for more complex games of imperfect information, like Magic or Netrunner. Although I suppose computers have a similar edge to the one they have in Starcraft, in that they can count cards and compute probabilities much faster and more reliably that humans.
I’m also interested in competitive card games because metagaming (i.e. having a good understanding of the dominant strategies and styles of play, how to beat the others with yours, etc.) is necessary to do well. How will we give the machines metagame knowledge? Having a metagame also complicates the definition of “fair” for human vs. computer play.
Indeed, even amateur bot authors report a significant edge:
This achieves an “idle” APM of about 500 and its battle APM is between 1000 and 2000 (as measured by SC2’s replay counter.)
I’m not into SC2, but I found some fans saying tournament-level humans operate around 150 APM with spikes into the 300-600 range.
I think metagaming is a distinction without a difference to AIs like Alpha Go.
I don’t think you’re correct, at least not for all games.
Consider this scenario:
I’m trying to build an AI for a rock-paper-scissors-lizard-spock tournament. An AI like Alpha Go will ultimately find the optimal strategy: pick uniformly at random. However humans suffer from biases, so a human who has studied their opponent’s habits may place better in the tournament than the AI.
In games like Go, this doesn’t seem to end up being particularly important. In games like Starcraft, I could be convinced that it’s not too important. In games like MtG, I think it’s quite important, at least until the AI successfully models deck construction as part of the game, which is qualitatively different than what Alpha Go has already done.
There’s metagame in Chess and Go, though they’re certainly much smaller than in MtG. The popularity of openings and responses to common gambits has changes over the years just like it does in MtG. It didn’t matter.
“Metagame” is a convenient distinction for people to talk about playing matches in the larger game of deckbuilding, but that doesn’t make it any less of a game. The task of analyzing and responding to “metagame” strategies is exactly the same as the task of playing the live game itself.
It’s going to be years before a state-of-the-art AI implementer turns to MtG, but I’m not going out on a limb when I say that humans will lose.
but I’m not going out on a limb when I say that humans will lose.
I agree that humans will lose eventually, but I think your original claim of “distinction without a difference” is too strong, because it asserts that humans will lose to AI that is largely of the same form as the AI now.
I would not be surprised (though I am not confident it is the case) if the success of Alpha Go’s architecture is limited in MtG and if competing with humans requires a different architecture.
computers have an inherent edge in the reflex-based/micro parts of the game
Tournament organizers might initially allow computers to have that edge, but eventually, to keep the games interesting, I expect they would cap the AI’s actions per minute to the level of a top human.
That’s what I recommend at first. It will help us assess what they can pull off given similar constraints to humans. I’m sure someone will complain that it’s basically cheating for humans because the computers could do better. In that case, I would be fair by limiting the computer to the size and energy consumption of the human brain but removing the actions per minute limit. The bot authors would ask to reinstate the APM limit instead.
I want to see it combine 3D object recognition
I want to see it combine 3D object recognition
Here’s an approach.
Your comment made me lookup the state of Arimaa, a game designed to be difficult for computers using then-standard techniques: TIL that computers beat humans at it in 2015 (source).