AI racks up insane high scores after finding bug in ancient video game

Video An AI bot managed to exploit a hidden bug in popular 1980s arcade game Q*bert to rack up impossibly high scores after being programmed using evolution strategy algorithms. Evolution strategies (ES) are interesting offer an alternative to the more traditional reinforcement learning (RL) methods that have been used to train machines to beat humans at Go or Chess. A paper published last week by a trio of researchers from the University of Freiburg, Germany, showed that in some cases the performance for ES can exceed that achieved with RL. It appears to flit between the platforms, jumping on the cubes in no particular order, causing the colours to change rapidly. Normally, the game should progress to the second round, but a bug causes the platforms to keep blinking and the agent can continue to bounce around and gain huge scores - close to one million points.