Bit of interesting news in the world of gaming from a while back; an AI beat a couple top StarCraft 2 players. Here is a link to get you started, but if you dig deeper there is a lot more context on how it happened, and how the AI evolved.
The main reason why the AI won, however, is that it was able to perform more action-per-minute (APM) than even a pro human player is capable of, and because of this, it was able to heavily out-macro during battles and win. Part of why this happened is that because in a game like Starcraft, the ‘strategy’ part isn’t nearly as deep or complex as the macro part, at least at the top end. Even a decent player will know the right build order, the right way to scout, and the right counters to what they see. Games at that level come down to who clicks better/faster when a battle breaks out, and its here that the AI was simply inhuman.
Which isn’t to downplay the accomplishments of the AI, it’s still impressive (and scary in a very Terminator-like way), but does raise the question about which games an AI could win at today, and which games I don’t believe it could. For starters, obviously an AI would dominate in a FPS, which is why aimbot hacks are banned in the first place. A slow, but zero RNG game like Chess is also one where an AI will win; it can simply calculate every possible combination, and always pick the best one (that Chess has a lot of said combinations is just a temporary issue, not a hard barrier).
I’d be very curious to see how such an AI would do in a game like League of Legends however. On the one hand, it would be very, very strong in things like perfect timing on last hits, and perfect coordination of skill usage and cooldown management. On the other hand, top players are already near-perfect at last hitting, and not that far off in CD management. But what makes LoL much different from SC is that LoL is less about APM, and more about situational decision making. How a team fight happens, and what happens in it, is a lot more chaotic than a SC battle, as there is simply more ‘stuff’ involved (more unique abilities, item builds, positioning, etc). I’m not saying the AI would lose, but I think it would certainly struggle more than in a game like SC.
I think the AI would get crushed in a game like Civilization. It would have zero APM advantage, but its calculating power like in chess would also be more limited as Civ has a lot of RNG in it, and how you react to that RNG is what really makes someone at great player. It’s also why in LoL when you face the in-game AI its decent, while the in-game AI in Civ has to get massive hard-coded bonuses to challenge a decent player, and even then by the mid/late game the AI is still pretty bad/dumb.
Fun topic to think about IMO, and one that I’m sure will evolve as AI development improves.
Still a long way to go before AI challenges any real human oppo imo.
Although Mr Musk recently said his own developed AI was too dangerous to be released as it could be used for nefarious purposes on things like social media so you can never say never.
Reading up on OpenAI in Dota 2 is also quite an interesting read. No success story headline for the media to jump on yet, but the development and the various strategies used by bots and humans, and them learning from each other is pretty fascinating. There are so many factors that can come into play during a MOBA match, and it’s quite enlightening to see which the bots are better at, and where human intelligence is still outmaneuring AI.
Also, let’s not forget that the initial strategy and methods that AI and other machine learning is set/decided on by another team of humans. That tends to lead to the AI being focused or trained or nurtured up in different ways. But it is nice to see emergent stuff emerge, where they end up hitting on hereby unexplored strats humans haven’t yet thought about, or utilizing areas where they are so much better than humans to eke out a win or put up a decent fight.
I mean, the fact that AI has conquered Go is a pretty big indication that there are no safe corners of human experience, IMO. Randomness is not going to be of particular value, because it affects both AI and humans equally. Besides, what is RNG going to do? Change the values of particular moves? The AI would adjust and run new simulations on all possible moves and pick the optimal solution.
The one question mark that remains would be metagame. The AI might play perfectly with character X, but if character Y has an advantage over X specifically, it’s possible that humans could essentially metagame the AI. That’s assuming that the AI can’t switch characters, and the advantage is not possible to overcome otherwise. Of course, who would actually want to play a game like that?
Re: Civ, the AI is bad because the designers are bad at AI. It has nothing to do with what’s possible with a few supercomputers networked together, or whatever Google was doing with AlphaGo.
Go is like Chess though, it has zero RNG. (The most interesting thing about the AI playing Go was that is completely changed the meta of that game, and significantly improved its play because it did so).
The reason RNG is a big deal when dealing with AI is that the one area AI is weak is in adapting, and RNG forces adaptation. I also don’t think a game that is more open-ended in gameplay/goals is an easy task for AI vs players, which is also why I think it would fail horribly in Civ. You can’t easily recalculate Civ every turn, because the next turn most of those calculations could end up being ‘wrong’ (you cut forests to boost building a wonder that is the cornerstone of your strategy, next turn someone else completes the wonder).
Civ wouldn’t be any more difficult for a good AI to stomp on human players than Starcraft. APM advantage is simply one of many advantages available to a PC.
Where AI truly shines in things like this, is its ability to influence the metagame. AlphaGo is actually a perfect example of this, as it has explored openings and move sequences that literally no professional human has explored given thousands of years. Taking that type of advance and applying it to any other strategy game is where AI shines, not just the inherent APM advantage.
I watched a few of the games, quite fascinating. The AI was actually APM-capped, but not sufficiently; should have gone further down by a couple of hundred. Also, not all APM is created equal. A lot of the APM for even a pro human player is not direct tactical orders but just moving the screen around, lightning-fast building queue updates, and other fluff to keep the flow going. The AI could execute deeply-considered moves, like godly stalker micro (btw, I think you switched micro and macro a couple of times in the post) several times a second, whereas a human player would need about half a second to process each such action. If more nuanced limitations had been placed on the AI, a little more in line with what we meatbags can mechanically execute, it would have been a more interesting contest.
Civ is an interesting question; I think that a self-teaching AI could very well be fine in a competitive multiplayer game after having developed a large enough set of heuristics. The thing about RNG is that it smooths out over a large number of dice rolls, and the AI would be better at keeping track of all the variables and sticking to the plan than a human player would. It goes without saying that it would wipe the floor with the existing Deity AI.
I can’t speak to LoL, but ask yourself this: how many discrete strategic rules do you actually apply when playing at a high level? I mean blink-recognising a particular kind of situation based on your experience as a skilled player and reacting to it in the optimal way? A hundred? A thousand? Hardly beyond reach.
On SC: The game I watched the APM wasn’t really capped, as it was executing at around 3k during the decisive battle if I remember correctly. That’s far beyond inhuman. That said, I think capping it to be ‘human’ might defeat the point of the AI, or at least highlights that ultimately SC comes down to who clicks faster once both sides know what to do (and ‘know what to do’ in SC is simpler than in LoL or Civ).
Civ: While RNG over many games does smooth out (sorta, Civ has some pretty extreme RNG in a lot of ways), per game it would still greatly impact the AI. The stuff it could calculate, like optimal city design taking all current and future upgrades into consideration, a decent player does nearly perfectly already anyway (and as mentioned above, if that calculation factors in a wonder it ends up losing, that could hurt the AI a lot more than a player).
LoL: The challenge in LoL, at the pro level, is that because you only control one champion, the required APM is much lower than in SC, while the way a critical team fight plays out is far more varied than a SC battle. I think that is why an AI would have a harder time vs SC. It’s less about the number of decisions to make, but more about getting the 1-2 decisions right when that surprise team-fight breaks out. I’d love to see how often the AI puts itself in the right positions, and how well it would react to a surprise bad position.
The games I watched were AlphaGo versus TeamLiquid’s ‘MaNa’ and a couple of their other players. APM was capped, per David Silver’s (AlphaGo teamleader) promise to keep the AI to things a human could do as mentioned in the CNN piece, but it was capped insufficiently to make enough of a difference.
Here’s the clarification from Silver’s and Vinyals’s AMA:
“I think this is a great point and something that we would like to clarify. We consulted with TLO and Blizzard about APMs, and also added a hard limit to APMs. In particular, we set a maximum of 600 APMs over 5 second periods, 400 over 15 second periods, 320 over 30 second periods, and 300 over 60 second period. If the agent issues more actions in such periods, we drop / ignore the actions. These were values taken from human statistics.”
I think you’re right in that ultimately SC2 comes down to micro (not least because the AI trained against itself, with no APM advantage over itself, and still came up with micro-reliant strategies) but that’s a pretty high and possibly irrelevant level of ultimate-ness. I disagree that SC2 macro is easy, though, even with the LotV changes. Pros make it look easy, but for anyone below masters league, learning to execute macro flawlessly will pay greater dividends than mastering micro. It’s quite easy to underestimate its difficulty, particularly under the mental pressure of having to manage complex battles at the same time.
Re: Civ – I haven’t played much VI (still a bit repelled by the graphics, should probably just get over that) but in V you couldn’t really rely on getting any wonder against humans (or even Deity AIs before industrial era) – and the AI would be fantastic at gauging the odds of all such gambles with the info at its disposal. Purely conjecturally, though, I’m guessing it would not bother with wonders and the subtler victory modes at all and optimise along the lines of some high production early-mid conquest zerg.
Re: LoL – I can’t speak the idiom of the game, but I meant not so much the number of decisions per match and more your overall library of experience that informs what you do on those few crucial decisions. Like “ah, yes, at this particular stage in the game, [teammate role x] is going to [location y] to do [thing z], the correct play for me is go to [location y1] to cover for them”, etc. A vast number of such heuristics (vaguely comparable with chess openings) can be taught to an AI, though I can see how communicating with humans (even through preset signals) and coexisting with imperfect human play might complicate things hideously. Come to think of it, I’d be really interested to see AlphaGo play some 2v2 or FFA games in SC2.
Civ wonders: If you pre-plan, you can usually b-line for a wonder and then, assuming the city is in the right spot, mass-clear to speed build it. You obviously invest a ton to do that, and if you fail its catastrophic, but most games it works (granted I don’t play Deity much). But again is that something the AI would take a chance on? And if not, even if you do try for a domination mid-game rush, being too far behind due to lack of wonders/expansion can make that nearly impossible, since combat is mostly about numbers vs skill/APM. Plus I’m not talking 1v1 Civ games, which indeed are much simpler.
Would an AI in LoL be playing one champion or all of them on a side? Probably to make it a real test you’d want something like 5 different AIs playing against 5 humans.
The AI would play one side, so 5 champs. I guess to make it ‘fair’ they would need to clone the AI and its hardware 5 times, and maybe have each AI practice/learn each specific role, which is what pro teams do.
Chess already has this occurrence when the World Chess Champion was defeated by IBM’s Deeper Blue back in 1996.