If your youth was misspent dodging 鈥済hosts鈥� in the 1980s video game Pac-Man, or the unofficial sequel Ms Pac-Man, you may be bemused to learn that a computer program has learned to play just as well, using tactics that it developed by itself.
In both the original game and Ms Pac-Man, the player must control a blob that munches its way around a grid, eating dots and fruit, while trying to avoid several pursuing ghosts. If the blob eats a special 鈥減ower鈥� dot, it can eat the ghosts for more points.
and at E枚tv枚s University in Budapest, Hungary, started by giving their Ms Pac-Man program a selection of possible scenarios, such as 鈥渋f ghost nearby鈥�, and possible actions, such as 鈥渕ove away鈥�.
Advertisement
The program randomly combined scenarios with actions to produce rules, and then played games using random combinations of those rules to deduce which ones work best.
Missed tactics
The program also set its own priorities, important for situations in which two rules conflict. The most important rule, it decided, was to avoid being eaten by ghosts, followed by pursuing any edible ghost. The next rule says that if all moves seem equally good, don鈥檛 turn back as you have already eaten the dots in that direction.
The resulting program narrowly outperformed average human players and Szita and L枚rincz say the work is part of broader strategy for analysing the weaknesses of AI compared to human intelligence when using video games.
鈥淕ames are interesting and challenging for human intelligence and therefore an ideal means to explore what artificial intelligence is still missing,鈥� the researchers say.
For example, although the program scored better than the average human player, it failed to evolve certain tactics that humans find useful, such as waiting for ghosts to approach before eating a power dot to maximise the potential effect of the dot.
鈥淣o such behaviour evolved in any of our experiments,鈥� the researchers say.
Journal reference:
