Pacman AI Projects

These projects were developed at UC Berkley to teach fundamental AI concepts, and are used at numerous universities, including Georgia Tech. In particular, the project on particle filters was very enjoyable and enlightening - once I got it working, I realized, somewhat naively, that surely this kind of thing (Bayes Nets/SMC) must surely be the future of optimization problem solving where analytical solutions are intractible or unknown.

In the particle filter Pacman project, Pacman is blind but has a kind of imprecise sonor-like senory ability, and instead of evading ghosts, he tries to eat them. His limited perceptions yield only a rough estimate of an area in the grid-like map where the ghosts might be found, with differening pobabilities of ghost locations being represented visually with different intensities of color. Pacman himself is pretty dumb - he just moves toward the closest of the brightest cell of each color, representing his strongest belief of where the ghost of that color is currently located.

When I need to explain multi-modal optimization to someone, I use the visualization of Pacman and his particle "sonar" tracking down ghosts as a visual aid. It's the best visualization of this concept I've come across.