If you can鈥檛 win, kill yourself or cheat. That鈥檚 the strategy invented by an artificial intelligence trained to play old Atari video games.
Patryk Chrabaszcz of the University of Freiburg, Germany and his colleagues created an AI to play eight Atari games, including the arcade classic Q*bert, in which players must navigate a strange orange character around a pyramid and dodge enemies. Rather than mastering the game like a human gamer might, their algorithm came up with two particularly unusual ways to play.
Advertisement
In the first, the AI gives up trying to win and instead baits an enemy into killing itself, before committing self-destruction too. This scores it just enough points to advance in the game.
In the second, the AI seems to have discovered a bug in the game that lets it cheat. The algorithm completes the first level, but rather than advancing to the second level as expected, the pyramid starts flashing and the AI鈥檚 score starts rising rapidly, reaching nearly a million points.
Algorithms that can play Atari games has been a growing area of interest since 2015, when Google DeepMind demonstrated an AI that could beat high scores set by top human players. The task isn鈥檛 idle fun 鈥� the algorithms are not told the rules of the game, but have to 鈥渨atch鈥� the pixels on screen to learn how to play. The idea is that game-playing algorithms could one day start learning about the real world as well.
DeepMind trained their algorithm with a technique called reinforcement learning, in which the system is rewarded for certain actions, like scoring points. Chrabaszcz and colleagues instead us an approach called an evolution strategy.
As the name suggests, an evolution strategy involves borrowing ideas about mutation and selection from biology. It works by starting with an initial strategy for playing the game, then at each point in time randomly tweaking or mutating the strategy to produce new ones. The system evaluates which of these 鈥渙ffspring鈥� strategies achieves the highest score and then further mutates the best-performing ones at the next time step. Over time, evolution ensures the best strategies, or at least the most successful ones, dominate.
Reference:听
