USM 2010 Mario AI Championship at the
Computational Inteligence in Games Conference, Copenhagen, Denmark, August 2010
Entries from USM student Slawomir Bojarski
Faculty advisor: Clare Bates Congdon

The 2010 Mario AI Championship included three different competitions, for artificial intelligence (AI) programs in Gameplay, Learning, and Level Generation. USM student Slawomir Bojarski entered his programs to compete in two of these competitions, winning first place in both the Gameplay competition and the Learning competition.

The goal of the Gameplay competition is for the AI program (often called an "agent") to play the game and win as many levels as possible. Every 42 msec, the agent must decide which keystrokes to press to move Mario left, stay, right, and up (jump), or down (duck). This competition evaluates the agent's ability to play the game in general.

The goal of the Learning competition is to play a random level 10,000 times and to have the highest score on the 10,001st attempt on the level. This competition evaluates the agent's ability to learn the hazards of a particular level.

The competition was held at multiple conferences during the Spring and Summer of 2010; results are posted online here.

Our approach is a evolutionary rule-based system, called REALM ( full paper available online here).

This run shows some wall jumps, piranha flower and bullet avoidance, yet got hurt from Spiky, but managed to devour a power-up flower!

This video shows how REALM agent deals with Dead Ends.

Longer story of sequence of successful runs. Watch out for incredible wall-jump double triplet, still avoiding the bullet just above the gap.

Attacking behavior learned by MarioAI agent.
Maintained by: Clare Bates Congdon (