Instructinator

Black and White – Learning AI

Description: Black and White is a game in which you play a god sent to a planet to do good or evil. The path you choose shapes the land and the inhabitants. You are given a creature of your choosing to act as your visible avatar in the world. But you have limited control over this creature and they must be controlled somewhat like a pet animal.

System:
In Black and White you are given a creature which you can train to work for you. You can teach them to play and work but the use of a praise/ punishment system. When you show the creature something an action, you can then access the interface for the praise/ punishment. Then you can either encourage the action or discourage it by either beating the creature or petting it.

Implication: This system could be used in any game where you have in-direct control of an avatar for prolonged periods of time. This is the foundation for a great system for teaching a program what behavior you would like to be performed, and what ones to avoid.

AI, of this nature is comprised of a set of available actions for a context. Say there is the context of a villager in front of the character. There is a list of actions to choose from. Either the creature can kill the villager, perform a food miracle and feed the villager, or pick the villager up and transport it to where it needs to go. If killing the villager has a lower rating then the probability that the creature will perform the killing action will decrease.

This way, simply making lists of actions with ratings in context we could perhaps have an enemy stalking the player, use ratings of actions in context to learn the players objectives and motives. Use that information for the AI character to formulate plans to track and thwart the player, or even move ahead of the player and lay traps where it knows the player is likely to fall for them. Think if Kefka had this AI in Final Fantasy 6. That would be a fun gameplay experience.
Jeff Mundee March 18, 2004