Two decades after Deep Blue conquered chess and a year after  AlphaGo took down go, there are still games that artificial  intelligence can’t beat.

  AI researchers like games because they offer complex,  concrete, and exciting challenges, which can unlock broader  applications. Take  it from AlphaGo creator Demis Hassabis: Games  “are useful as a testbed, a platform for trying to  write our algorithmic ideas and testing out how far they scale  and how well they do and it’s just a very efficient way of doing  that. Ultimately we want to apply this to big real-world  problems.”

How close is AI to beating different games? We  asked a couple of AI experts for an update.

  “StarCraft” is the next big target

  After Google’s AlphaGo   defeated grandmaster Lee Se-dol in go — that deceptively  simple game of black and white stones with an immense number of  possibilities — Hassabis called “Starcraft” “the  next step up.”

  Google engineer Jeff Dean, who leads the Google Brain  project, went further: “‘StarCraft’  is our likely next target.”

  “Starcraft,” released in 1998 by Blizzard Entertainment, is a  real-time strategy game where players build a military base, mine  resources, and attack other bases.

  It has become a focus for AI researchers for a few reasons: it’s  highly complex, with players making high-level strategic  decisions while controlling hundreds of agents and making  countless quick decisions. It’s popular, as the first major  esports game and practically the national sport of South Korea in  the 2000s. It also has a relatively big research community with  annual AI competitions dating back to 2010, a publicly  accessible programming interface, even support from  Blizzard.

  Despite all this attention, the best “StarCraft” bots can easily  be beaten by top humans. The bots tend to be good at  low-level unit interactions but bad at high-level strategy,  especially when it comes to adapting that strategy on the fly.  They may also make dumb mistakes.

  “There is progress, but it’s maybe a little bit underwhelming,”  says Julian Togelius, the co-director of the NYU Game  Innovation Lab.

Human player Djem5 destroyed bot tscmoo in this 2015  match after finding an undefended base.AIIDE

  This has a lot to do with the type of people building “StarCraft”  bots. Brilliant though they may be, they tend to be solo  hobbiests who have nowhere near the resources of large companies.

  That could be changing. In addition to Google’s hinted interest,  Facebook and Microsoft researchers have published papers  involving “StarCraft.”

  “With the increased attention that “StarCraft” AI is getting from  large companies, if they do end up spending time and money on  this problem, this will accelerate progress dramatically,”  says David Churchill, a professor at the Memorial University  of Newfoundland who runs the “StarCraft” AI tournament.

  Even if big companies do get involved, beating “StarCraft”  will require breakthroughs we haven’t seen before. Deep learning  — the seemingly miraculous technique of feeding computers  extremely large sets of data and letting them find patterns  — was central to Google solving “Breakout” and other simple games; it  helped Facebook understand “StarCraft”  combat; and it played a significant role in cracking  go. And yet …

  “I have a feeling this is not going to work so well for  “StarCraft,” Togelius says. “You need planning to play this  complicated game, and these algorithms, deep nets, are not very  suited to it.”

  Some games are hard, some are easy

  While “StarCraft” is the white whale, plenty of videos games  present their own AI challenge.

  Turn-based strategy games like “Civilization” require similar  strategic decisions, often with more long-term planning, but less  rapid unit control.

  “We don’t have very strong playing ‘Civilization’ agents but they  haven’t done that much work into it,” Togelius says.

  “In my opinion, a game like ‘Civilization’ would be strictly  easier for AI to beat than a game like ‘StarCraft,'” Churchill  says.

        Can  bots play “Civilization” better than humans? Ominous  question.                  2K  Games/Firaxis       

  What about “League of Legends,” the popular game where players  control single heroes in complex team battles?

  “I have talked to certain people who think that it would be  really easy to make a ‘League of Legends’ AI that could beat the  world champions, but I don’t necessarily agree,” Churchill says.

  “It would be easier than ‘StarCraft’ in the sense that you don’t  have so many different levels of control, but you’d still have  the issue of teamwork, which is very complex,” Togelius says.

        Can bots win at “League of Legends”?        YouTube/Riot

  “FIFA” and other sports games could be difficult, too. As with  “StarCraft,” they require control of an extremely high number of  variables.

  “You have 11 players, and if you say each player can do one of 4  different things, then your number of possible actions is four to  the power of 11 — which is about 4 million,” Togelius says. For  comparison, he notes, chess players choose from around 35  possible moves every term, go players from around 300.

  For first-person shooters like “Call of Duty” and “Halo,” being  able to program a bot with perfect aim is a huge advantage.

  “Some people think if you have perfect aim, then AI will never  lose,” Churchill says. “But others say there are strategic  elements of the game that can overcome perfect aim.”

  Fighting games like “Street Fighter” are relatively easy for  bots.

  “Unfortunately I think too much of the challenge is in having  fast reactions, and you can have arbitrarily fast reactions if  you are a computer,” Togelius says.

  Platformers are relatively easy too.

  For instance: “’Super Mario Bros’ is essentially solved, at least  for linear levels,” Togelius wrote in an email.

        This 2009 AI was able to  crush “Super Mario Bros” (red lines show paths being  evaluated).                  Robin  Baumgartner       

  Role-playing games can be tricky.

  “Skyrim” requires “understanding narrative” and “a wide variety  of cognitive skills,” Togelius writes online.

  As for card games — remember those? — they still pose some  interesting questions. Two-person limit Texas Hold’em has been  solved,  but group games and unlimited betting options are still a  problem. In Bridge, bots haven’t  mastered deceptive play, bidding, or reading opponents  and still can’t beat the world champions.

  The playful linguistics of crossword puzzles are a challenge as well.

  And we’re not even getting into   robotics (Cristiano Ronaldo is way better than robots at  soccer).

  The ultimate challenge is general game AI

  A few years ago, Togelius noticed a problem with video game AI  competitions. People inevitably start programming bots with  extensive human-knowledge about how to beat a game, rather than  focusing on true artificial intelligence.

  “People submit more and more domain-specific bots — in the  car-racing competition you had people submitting bots that  started hard-coding elements of tracks and cars and so on — and  the actual AI part gets relegated to smaller and smaller roles,”  Togelius says.

  It was this frustration that led him to start the General Video  Game AI (GVGAI) competition in 2014. In this competition, players  submit bots that compete in an unknown set of ten simple games.  To win, the bots have to be flexible and adaptable — something  closer to humans.

  “General intelligence is not the capacity to do one thing, it’s  the capacity to do almost anything you’re faced with,” Togelius  said.

  For now, the bots in GVGAI aren’t great.

  “On most games, [they are] not human-level,” Togelius says. “On  fairly simple shooters, such as ‘Space Invaders, you have bots  that are better than humans, but on things that require long-term  planning, they are often considerably under human level.”

  Give them time. Togelius says interest in the competition is  growing fast. Notably, one of its sponsors is Deepmind,  the Google subsidiary behind AlphaGo. 


              “Space Invaders” is easy  for bots.        Midway           

  Human-machine teams could be the future

  As humans are surpassed in game after game, we can hold onto one  reassuring fact: centaurs could do better than humans or  computers.

  Centaurs, a term for human-machine pairings taken from the  mythological creature that was half-human-half-horse, have  already proven more effective than humans or computers at chess  (at  least when given enough time, at least for now).

  “The human uses his or her intuition and ideas about what to do  and ideas about long-term strategies and uses the computer to  verify the various things and to do simulations,” Togelius says.

  A similar pattern might hold true in “StarCraft”— and beyond.

  “I could very much imagine a future where we have centaur teams  in ‘StarCraft,’ where you have a human making high-level  decisions and various AI agents at different levels executive  them,” Togelius says. “That is also how I imagine the place AI  technologies will more and more take in our society.”

  Other people have drawn similar conclusions. Togelius pointed to  roboticist Rodney Brooks, who predicts in the classic book “Flesh  and Machines” that future humans “will have the best that  machineness has to offer, but we will also have our bioheritage  to augment whatever level of machine technology we have so far  developed.”