January 22nd, 2008 by Arjan Olsder Posted in Other news | No Comments »
A bit off, but this morning I read an interesting bit of news regarding
software agents that learned how to play the classic Ms. Pac-Man.
The software agents where developed by two employees (Istvan Szita and Andras Lorincz) of the Hungarian Eotvos University. The agents use a technology called reinforcement learning during the Ms. Pac-Man game (Atari 1986).
The choice for Ms. Pac-Man was important to the developers as unlike earlyer versions, the ghosts in the game don’t follow the same route. This way, the software agents can’t predict where the opponents will go. This way, the software has to make predictions based on observations.
The employees also tested the agents against human players. They compared the results after a total of 50 matches. The software agents where slightly better then the humans. In another experiment with the same agents, but without the ‘self learning’ mechanism, the human players excelled.
There was a big difference in play style though. The software agents based their routes purely on observations of the ghosts whereas the humans tried to make predictions of where the ghost would pop-up. This is something that was too advanced for the software agents to predict. This could mean that if the software agents where further perfected, they could beat the human players with a bigger difference.
