Based on Peter’s descriptions of the learning mechanisms, it seems a decision tree is still used for understanding objects (like in Black & White). When you give feedback to your dog, it gets an extra data sample to build the tree.

In Fable 2, you can train your dog in the typical way, by rewarding it or punishing it. For example, if your dog likes pooping in the neighbor’s flowers, you could correct that with the basic YES, or NO emotions of your avatar (depending on whether you like your neighbor or not). The same goes for the dog playing with his ball, or burying it.