(Attractiveness vs. Efficiency)

2 Background

Computer animation techniques which complement traditional animated scripting with autonomous agents have made possible complex life-like systems composed of many distributed elements (Reynolds, 1987). Physically-based modeling techniques and virtual motor control systems inspired by real animals are used to automate many of the subtle, hard-to-design nuances of animal motion (Badler, 1991). In task-level animation, (Zeltzer, 1991), and the space-time constraints paradigm, (Witkin and Kass, 1988), these techniques allow an animator to direct a character on a higher level.

The idea to use the genetic algorithm (GA) for automation of animated motion follows naturally. One application in using the GA for evolution of goal-directed motion in physically-based animated figures includes evolving stimulus-response mechanisms for locomotion (Ngo and Marks, 1993). While the work described in this paper bears a resemblance to the virtual creatures of Sims (1994a), it continues a previous line of explorations, using a GA for optimizing locomotion in physically-based figures. The first example is a 2-legged walking figure (Ventrella, 1990) which evolves locomotion through pursuit of food, continuing with an unpublished 1992 project in evolving morphology and locomotion in 2D swimming creatures, and later, including evolution of 3D morphology as well as anatomy for locomotion (Ventrella, 94). Sims (1994a, 1994b) has developed techniques for the evolution of morphology and locomotion most comprehensively and impressively, using the genetic programming paradigm (Koza, 1992), and includes extensive 3D physical modeling. A holistic model of fish locomotion, with perception, learning, and group behaviors, has been developed by Terzopoulos, Tu, and Grzeszczuk (1994), which generates beautifully realistic animations.

Evolutionary modeling of situated organisms which reproduce spontaneously takes evolutionary modeling a step closer to nature (Ray, 1991). "Electronic primordial soups" involving spatiality, such as Yaeger's Polyworld (1994), demonstrate artificial ecosystems in which mating, eating, learning, and even social behaviors, evolve within the simulated world.

In a prior paper (Ventrella, 1996), it was demonstrated that swimming skills in physically-based figures could evolve without the use of an explicit fitness function, by delegating reproductive freedom to the organisms in the simulation such that locomotive skill emerged through competition for mates and food. (Throughout this paper I refer to this work as the previous simulation). A commercial product was derived from this work (RSG, 1997) which enables users to manipulate the organisms and conduct experiments.

Ijspeert, Hallam, and Willshaw (1997) have developed artificial neural controllers which evolve through a GA for optimizing swimming locomotion in simulated lampreys, which can produce complex oscillations for undulating-style locomotion.

Todd and Miller (1991) developed a model which demonstrates how the forces of sexual selection can drive a population to have arbitrary phenotypic features, above and beyond the features resulting from natural selection.

The current project is motivated in part by Todd and Miller, and explores mate preference, building upon the previous simulation. As an extension of the previous simulation, the current simulation includes:

1) a more comprehensive physical model with a larger phenotype space
2) a set of pre-defined mate preference criteria
3) a means of measuring the effects of mate preference on evolution of locomotion skills

3 Fitness

(go to beginning of document)