Well, there’s plenty of complex (and not-so-complex) games with rubbish AIs, but that’s because making AIs is hard, especially for a game that’s not been solved, and often was still evolving and being re-balanced 5 minutes before it shipped. I’m not sure a procedurally generated, multifocus game like *Civ *could even be solved - certainly the debates over what’s proper play keep raging on. So I wouldn’t hold my breath for a perfect AI for it, even if someone did try and make it as a coding challenge.
Checkers, chess, reversi… - simple games with simple rules and limited moves yet emergent issues complex enough that they can’t reliably be solved for. It’s already very difficult to make good AIs for those games, even with the sheer brute force of Deep Blue and the like to crunch future plans N moves ahead (well, checkers might be easier, admittedly). Go, as mentionned, is pretty much unsolvable even with brute force since the “correct” play has as much to do with philosophy than anything and the board’s state flows a lot.
Civilization, in many ways, is N-dimensional Go with X players at the same table. Of course the AI’s gonna suck at it. Most *people *suck at it.
So let it roll a dice on whether to bluff or not on a given hand/situation. Unpredictable unpredictability.
Civilization is inherently not solvable because it’s not deterministic. No matter how good your plans are, there’s always a chance that a stone-age spearman will defeat your Abrams tank. In principle, though, with enough brute force computing power, you could calculate the probability of winning from any action with optimal opponent response, and thus in practice “solve” it.
Go is definitely solvable, and again, a brute force method would in principle do it. The “philosophy” of correct play is really just a bunch of heuristics that have been developed because the brute-force solution is too impractical.
I wonder how well AI fares with the various collectible card games (Pokemon, Magic: the Gathering, Yu-Gi-Oh!, etc). Especially in a “random draft” style situation where the decks aren’t predetermined and so the AI has to make a workable deck out of a bunch of random cards, shuffle it and play a human player. I know there’s computer versions of some of these games but I don’t think they have the same randomization you find in a real world random draft.
Diplomacy. When the world Dip champ is a computer, I’ll acknowledge that the goal of ‘true AI’ has been achieved, or at least close enough as to make no difference.
I’m surprised that a pair of computers haven’t already been able to outplay the best humans.
I was just going to say MtG. I had a coworker a few jobs ago (this guy) who was very involved in Go AI, and I claimed that writing an expert-level Magic AI would be harder than Go.
That means, to me, that the AI has to be able to look at a new magic card that it’s never heard of or seen before, and figure out how good that card is, how best to play that card, how to build decks with that card in them, how to play against that card, etc. That is a HAAAAAARD problem.
Max, I don’t know if you ever played the old Shandalar MtG game from Microprose, but that was with a very limited card pool, and the AI was godawful. It would giant growth your creatures, fail to attack when your board was empty, etc. Just miserable.
I can see that. Often, decks are made or broke by specific combos of two or even three cards, and the number of card combos is combinatorically larger than the number of cards. Plus, even getting the right cards out doesn’t guarantee that you’ll see how they can most effectively interact, since many combos depend on subtle nuances of timing and the like. And then you have to evaluate how consistently you can pull that combo off, and how good the cards are without the combos, and how many other combos you can make out of those cards.
True, although that doesn’t really prove much. I suspect that if we limit the card pool and ignore deck building it would be not particularly hard to build an AI that was damn good purely at playing with an already-created deck out of limited cardpool. HOW good I dunno, but certainly WAAAAAAY better than the comically awful Shandalar AI.
There are lots of games where the AI can only win consistently by cheating. Civ is a great example, all varieties. The only AI that gives a challenge is where everything is easier for the AI (bonus resources, fewer requirements for buildings or upgrades, etc.)
Also for example, first person shooters. The only chance the AI has in your average first person shooter is to always know where the players are at all times, and what weaponry they’re carrying. If they experience the same “fog of war” the players experience (i.e. that you have no idea where the enemy is unless you see or hear them), AI bots are trivial to defeat.
A human will annihilate a computer player at any relatively complex strategy game if the computer is playing by the same rules. That has always been true of ALL strategy games, from Civilization to Empire.
Civilization (of any version) is many orders of magnitude more complex than chess, so a computer’s strengths of brute power are vastly overwhelmed by a human’s strengths in developing heuristics.
I nearly quit playing the computer version of Ticket to Ride because the AI is so dumb, dumb, dumb. The game designers definitely could’ve done a better job, but as you said, it’s virtually impossible for any strategy game to accurately emulate a human.
This isn’t entirely true. Starcraft II can be played fairly well by AI. Not Blizzard’s AI, mind you, but some AIs that researchers have made. One could argue they don’t play by exactly the same rules, since the AI doesn’t have a keyboard, mouse, and screen, but when it comes down to it, these research AIs don’t outright cheat (they don’t have map vision or anything). Largely the way they win is by being extremely good multitaskers and using exceptional micro and macro skills to optimize their play to a degree that makes up for their decent, but sub-human ability to strategize.
I’ve played the most recent game (Duels of the Planeswalkers 2014) and the AI is fairly competent. Not ingenious by any stratch and it’s pretty easy to read it when it’s got a combat trick or a counterspell in hand, but it’s not the Shandalar trainwreck. It probably helps that its decks are all premade, most are single colour and all are fairly straightforward to play (it does noticeably worse on the few “gimmick” decks, especially the Sliver deck. It just doesn’t know to keep the important booster slivers out of harm’s way even if you’re telegraphing your own tricks).
No draft though - there’s a draft format in the game, and the human player gets to rip open booster packs and build up his deck with the luck of the draw, but then they’re pitted against regular pre-made decks too.
All the people saying “Go” are kinda outdated. Go is being played by computers at pro levels right now - a professional 9dan has lost to a computer playing Go albeit with 4 stone handicap.
Zen matches against Ohashi Hirofumi and Takemiya Masaki were announced in February 2012.[21] On March 17, 2012 Zen beats Takemiya 9p at 5 stones by eleven points followed by a stunning twenty point win at a 4 stone handicap. Takemiya remarked “I had no idea that computer go had come this far.” - Computer Go - Wikipedia
In this case though we’re getting beyond “strategy game” and into just an outright video game. Starcraft is real time and so speed and reaction time can be as important as strategizing. If you converted it into a turn based game, the computer would lose to any reasonably experienced player.
On this subject, I always like to quote something I once read in fortune cookie, or on a napkin at a bar, or some such place:
“Both a computer and a person can win a great game. But only a person can enjoy it.”
Yeah, I’d suspect that you could program an AI to do well with a single tight deck, especially if you’re limiting the deck choices for the opponent. I’d consider that a pretty narrow definition of “playing Magic” though and wouldn’t be especially impressed by it. Unlike board games like chess or even games like poker, too much of Magic involves the ability to build something from a vast assortment of (potentially randomly assigned) cards and I can’t imagine the AI holding up under a draft tournament situation, for example. Even if you’re not talking draft, if the AI is having a human-constructed deck handed to it with specific instructions how to play it, it’s not really the AI doing 75% of the job (deck construction).
