The animation of artificial life in motion is a difficult subject because a number of complicated motions are involved by creature itself or by its interacting with others.  Moreover, there are no standard quantitative measurements for evaluating the performance of an animation.  Mostly the evaluation was based on visual observation, which could be very subjective and could vary from person to person. A lot of animation has been done by applying dynamic physical principles.  However, only physical principles are not good enough if there are several creatures living in a complex environment.  For instance, how could one creature know whether to take the action of looking for food or running away from its predators at certain moments, which can not be decided by physics-based model.  Hence it is important and necessary to explore other possible computation algorithms besides the physics-based model so that complicated motions involved could be simulated.  One step further, automated motion learning ability could be achieved by virtual creatures.

        The objective of the study was to develop an automated motion learning procedure for virtual creatures, which controlled and determined their dynamic movements of each creature and the interaction of creatures to each other and to the environment.  Particularly:
 
        1. To automatically evolve the creature's genes through genetic algorithms (GA),
        2. To develop neural networks (NN) which could decide the extent of an action taking place, and
        3. To derive fuzzy cognitive maps (FCM) which could be used to coordinate between any two
            distinct actions taken by creatures.

        Ultimately, the creatures should be capable of learning which and how the motions and interactions take place automatically and naturally.  Therefore, a virtual environment with high level of animation could be guaranteed.

        An evolving virtual environment (or ecosystem) was developed in this study by computer graphic and animation.  Artificial agents including two kinds of creatures (one fish and four bugs) and five plants (fire) dwell in the environment, with fish living by eating bugs while bugs by feeding on plants.  In this study, the creatures were modeled as symbots with sensors (eyes) which made them see both their food and their predators.  The sensors were connected to wheels via a perceptron neural network and functioned as inputs of the networks.  The outputs of the networks determined the actions of the wheels being taken by the creatures, such as hunting for food or evading from their predators.  The weights of the neural networks between the inputs and outputs were derived from genetic algorithm evolution.  Each creature had its own neural networks and some creature had two networks associated with it, with one performing food-finding task while the other one performing predator-evading task.  Fuzzy cognitive maps of the system were designed by relating the causal effect between states and were used to govern the overall function of the virtual environment by coordinating among agents and within an agent.  For instance, by switching between the two different networks, fuzzy cognitive maps provided the transitions from food-finding action to predator-evading action or the other way around.

        The experiment was conducted to determine some appropriate parameter settings for the ecosystem by trial and error.  The results of the initial study demonstrated an acceptable and a close-to-nature virtual environment was established with the existing settings.  The goals of achieving motion learning abilities were in some degree reached.  However, the reasons behind those settings were not investigated yet in the experiment.  It is needed to go beyond those superficial values and explore the theoretical background for those settings.  By modifying the parameters such as the number of the fish, the FCM edge connection values, and so on, observing any changes occurring in the ecosystem, and relate the cause and the corresponding results, hopefully  some theoretical algorithms could be derived.  If such algorithms are found, it would be more efficient to create a virtual environment with the aid of new findings.  Additionally, some improvements in the motion control of the creatures could be expected.