IF THERE’S ONE THING artificial intelligence (AI) loves it’s gaming, and now it has uncovered a novel loophole to beat the classic Atari game Q*bert.
But first, for those who aren’t aged gamers from yesteryear, Q*bert involved players jumping around a pyramid constructed of cubes, whereby each cube the land on changes colour and the objective is to change all the cubes’ colours while avoiding the game’s enemies.
Researchers Patryk Chrabaszcz, Ilya Loshchilov and Frank Hutter from the University of Freiburg had several AI programmes gaming away on an updated version of the Atari game to test their work on “evolutionary algorithms”, the BBC reported.
These work by generating lots of algorithms and putting them to test; the ones that perform the best and are modified and then put through more testing to see of the improve or get worse, essentially evolving the code.
Using such algorithms on Q*bert, the researchers were able to create an AI that through trial and error fond holes in Q*bert’s code and allowed it to score vast amounts of points
The AI discovered two methods to get a super high score that it’s believed no human players have figured out.
One involved making the character Q*bert commit suicide over and over, prompting the game to record enough points for giving the player another life while keeping the points total building.
The other more prominent method uncovered by the AI was a bug that is triggered when a player seems to jump around at random over the pyramid, causing cubes to blink and reward the player with a load of points as Q*bert continues to hop around on cubes.
According to the BBC, Warren Davis – who worked on the original version of Q*bert for arcade machines – noted that the bug could be a feature of the updated game code and doesn’t seem to exist in the original code.
“This certainly doesn’t look right, but I don’t think you’d see the same behaviour in the arcade version,” he said, noting he was not familiar with the ported game’s code.
While the likes of DeepMind’s AlphaGo have beaten human champions at games, in the case of the Q*bert playing AI it has shown that it can dig up exploits to well exploit rather than be constrained by reinforcement training to learn the best methods to play a game within its rules.
Not only does this show that the use of evolutionary algorithms could be a solid alternative to training AIs using deep reinforcement learning, it could also demonstrate how AI systems could be used to dig out bugs and holes in other software that humans and traditional testing techniques could overlook. µ
Source : Inquirer