The most appropriate action can be evaluated with respect to the goal of action selection. Animals in the wild often face hazardous situations. The appropriate choice of actions is crucial to their long term survival. Therefore, according to Dawkins Dawkins89, the ultimate goal of action selection for animals is to choose successive actions (or behaviors) so as to maximize the number of copies of its genes in future generations. That is to say, the ultimate goal is to survive and to reproduce. This goal breaks down to more immediate, day to day behavioral needs. For a robot, the action selection problem entails maintaining the safety of the robot while pursuing the successful completion of the tasks it has been assigned. For a virtual animal (or an autonomous animated creature in general), the goal of action selection is to achieve satisfactory behavioral realism. Since a mathematical representation of the above goals would be extremely complex, we may not be able to build action selection mechanisms via numerical methods, such as optimization. This makes the problem of deriving action selection mechanisms a design problem and hence the main issue is to come up with the corresponding design criteria [Werner1994, Maes1991a, Tyrrell1992].
Although the goal of action selection in a real animal seems different from that in a virtual animal, they are in fact similar. The action selection mechanisms in animals allow them to take appropriate actions in the face of uncertainty. This is what we refer to as rational, adaptive behavior. When we say the behavior of a virtual animal looks realistic, we generally mean that the behavior it exhibits makes sense. Like a real animal, the virtual animal ``knows'' to avoid hazards, to exploit opportunities and to reasonably allocate resources, etc. Our approach to achieving such realism in animation involves identifying the important principles by which animals select actions, especially priorities between different behaviors, and employing these principles as the design criteria for building action selection mechanisms in virtual animals.
|Xiaoyuan Tu||January 1996|