GONE are the days when video gaming was a private pursuit. Gaming services such as Microsoft’s not only connect players in living rooms the world over, they can also record every move each gamer makes. Academic researchers are learning to use information mined from this mountain of data to build more stimulating games – and commercial games designers are beginning to take notice.
“All of the big games publishers are getting into data mining,” says of the Center for Games Research at the IT University of Copenhagen, Denmark. “They’re talking to universities, even hiring researchers to work on some of these huge data sets.”
The trend is all the more remarkable because games designers are usually reluctant to collaborate with academics. Togelius says that designers find most aspects of academic games research, such as artificial intelligence (AI), too esoteric to use as part of the development process. Using data mining to study how gamers play existing titles, though, can give developers instant rewards, such as identifying points in a game where players are likely to become frustrated or bored. The insights could help to tailor future releases to make them more satisfying.
Advertisement
There is a problem, however. The data sets are so dauntingly complex that analysing them can defeat even the most skilful and experienced games designer. But here smart software developed by academic researchers can step in to help uncover patterns that are hidden from humans.
The easiest way to treat the data is to look for direct correlations – perhaps a large number of players losing a life at a particular spot. But other patterns are only revealed by delving deeper into the data. To do this, researchers turn to algorithms similar to those used by banks to pick out fraudulent behaviour from a mass of legitimate transactions. “Machine-learning algorithms are great at finding patterns,” says at the University of California, Santa Cruz.
At the (CIG 2010) in Copenhagen this week, Togelius, and are on data mined from 10,000 Xbox Live gamers as they played . “It turns out that we can rather accurately predict whether or not a player will finish the game by just looking at a few features of their game play,” says Togelius.
For each gamer, Togelius and his team identified several features of play, such as how much time they spent in a particular room in the first level, and the number of rewards they collected. The team then fed the data through software containing a suite of prediction and classification algorithms to produce their final predictions.
Researchers are also using data mining to improve computer-controlled characters within the game, which helps them react appropriately to the range of different strategies that players can adopt in today’s complex games. Programming these characters to react sensibly to different tactics is labour-intensive. “That’s motivating the need to begin automating the process,” Weber says.
To explore what can be done with artificial agents, Weber has cribbed data from the mother of all computer strategy games, StarCraft. Released by Blizzard Entertainment in 1998, StarCraft pits three alien races against each other in a bid for dominance. Over the years, its devoted legion of fans have fine-tuned their strategies to levels of sophistication on a par with those of a chess grandmaster, Weber says.
To help newcomers get into the game, community websites host replays of previous games between the top players. The idea is that novices study these games to learn which strategies are winners, but AI researchers are now using this resource to enable their software to do the same.
Weber has used this approach to create a robot player called EISBot. He downloaded thousands of replays, and used machine-learning algorithms to identify patterns in the data that helped predict how games would unfold. That knowledge was then encoded into EISBot. After only a few minutes of game play, EISBot can predict an opponent’s strategy with 70 per cent accuracy at least 2 minutes before it is executed – an advantage in a real-time game.
“The goal is to create a bot that is a challenging opponent, but not so good it beats you every time”
If anything, robots like EISBot play games too well to be incorporated into commercial games. “Gamers will expect more and more realistic behaviour from the characters in games,” says Johan Pfannenstill, a lead programmer at based in Malmö, Sweden. “It is very important to try to meet those expectations.” Make characters too smart and they “jeopardise the intended experience for the player”, he says.
Pfannenstill might be more comfortable with the approach being followed by at Edith Cowan University in Perth, Australia. He is also using data mining – this time from the first-person shooter game Unreal Tournament 2004 (UT2004) – to make bots behave more like human players. His studies suggest gamers prefer these opponents.
Hingston’s goal is to make the bots just intelligent enough to pass as human rather than so intelligent that they take control of the game. So he has begun using UT2004 as the backdrop for a of the Turing test. For this, he sets up a UT2004 environment in which both humans and bots are playing, and asks the human players to say which they think is which.
As 91av went to press, the latest round of Hingston’s Turing test for bots competition was in full swing at CIG 2010. Hingston thinks he knows what characteristics will define the victor. The bots “need to strike a balance between appearing superhuman and too stupid to be human”. Despite Pfannenstill’s concerns, most data-mined AI bots excel at the latter but not yet at the former.
