(Artificial Emotion - 2)
As long as there is a clear set of critera which tell the machine what are supposed to be "good emotions" and what are supposed to be "bad emotions", it will be able to figure out what problems need to be solved in order to maximize its chance of "feeling good instead of feeling bad".
We could imagine a hypothetical space called "emotion space", in which each point can be defined as an individual emotion (similar to an "idea" in Plato's universe). Emotions which make the machine "feel good" are a cluster of points in the emotion space, and emotions which make the machine "feel bad" are yet another cluster of points in the emotion space.
And the rest is quite straightforward. The two aforementioned clusters are essentially just "points you must move toward" and "points you must move away from", respectively. Mechanically speaking, they are functional equivalents of positive and negative electric charges in empty space, where the observer is an electron and therefore inclined to be pushed by negative charges and be pulled by positive charges.
The AI mind, whose current mood can be expressed as a single point (electron) in the emotion space, can then simply issue a pathfinding problem which will let it find out the shortest emotional path that leads the point to the "center of mass" of the cluster of positive emotions.