(Sexual Swimmers)

1 Introduction

The majority of species on earth have evolved skills for locomotion in water, and the variation in morphology among species is matched by a great variety of swimming styles. For most species, locomotion is necessary for getting to food and, for many, to avoid becoming someone else's food. Locomotion also plays a role in the reproductive lives of most species. In order for an individual to mate with another of its species, it needs to approach this individual to be close enough - obviously - to mate. The simulation described in this paper bases its ability to generate emergent locomotion and morphology from these two simple elements of nature - food and sex.

This paper describes a simulation which drives an artificial life movie - a movie with no script. The actors are a variety of 2D figures, called swimmers, which inhabit a virtual pond. The swimmers' bodies consist of interconnected line segments of many colors, with movable joints. The motions of body parts can exert forces against the water and thus potentially accelerate a swimmer through the water. In this world, swimmers eat, reproduce, and die. Swimmers who are able to swim to food can replenish their energy. Those who cannot, die of starvation. Swimmers who are both able to eat and able to swim to a chosen mate, reproduce. Those who cannot mate, do not pass on their genes to future swimmers before dying. Evolution of improved swimming results. On the local scale (individual swimming styles) as well as the global scale (group population dynamics), emergent phenomena can be observed.

Sexual selection is modeled here as well - each swimmer has a genetically inherited favorite color in potential mates, which influences which swimmer it chooses from those within its view. These preferences are randomized at initialization, along with all the other genes. Eventually, in maturing populations, swimmers' color preferences begin to correspond with their actual body colors, often catalyzing distinct groups to emerge, in which case, the population becomes sympatric, with groups rarely interbreeding, particularly among the "purest" races, in which body color is homogeneous.

The use of color in modeling mate preference is not for aesthetics (not ours, at least - perhaps the swimmers'), but as a readily-visualized way of studying the effects of mate preference.

1.1 Autonomous Reproduction

Operators derived from the genetic algorithm (GA) are used in this simulation, but not in the conventional manner: in this case there are no discrete generations, and there is no explicit use of a fitness function for assigning fitness values to individuals in a population, thereby determining the rate at which they can reproduce. Instead, creatures who are able to swim towards a desired mate automatically reproduce, by virtue of the fact that they are able to swim to their goal. The fitness landscape is continually changing, as in the natural world. Fitness, then, is equivalent to reproduction, which, in this world, is equivalent to coming in contact with a desired mate.

At no time is any "outside assistance" used in the simulation to help encourage the emergence of optimized swimming - other than in the initial distribution of swimmers and food, and a set of swimmer perception and behavior settings. There is nothing but the situation at hand to determine how the population will evolve. In many simulation runs, the entire population of swimmers dies off before any significant results occur. This is the occasional price to pay for a decidedly hands off approach.

This approach also offers no means of exploiting a good solution when it happens somewhere in the pond. An Olympic individual can emerge within the pond, but if it has an inappropriate preference for mates, is not desired by other mates, or is born in a low swimmer or food populated area, it could die without ever having reproduced. This simulation runs on a mixture of skill and luck.

This work can be seen as an attempt to extend recent methods in generating emergent morphology and motor control in virtual creatures using GA's. In this simulation, the changing environmental conditions (and not explicitly a "Creator's" objective function) determine emerging behaviors, which are transformed by (and likewise transform) the environment. The emphasis, then, is not on simply optimizing behavior within a population, but in modeling and studying the heterogeneous outcome of a population of creatures living their lives within an ever-changing, varying ecosystem.

1.2 Real-time Animation

The simulation described here has an interactive component - the design of which has been a useful investment for development and analysis. One can view the running of a simulation using an interactive "microscope" which can zoom in and out, or pan across the pond. With this microscope, different parts of the simulation can be studied - one can watch a single swimmer up close; view a genetically related group of swimmers in one area of the pond; or view the entire pond to witness large-scale dynamics.

Populations of about 200 or less can be watched at real-time computer animation rates (10 or more frames per second) on an SGI Indigo2 computer. In experiments, the population count is usually initialized to 1000 - which is too high for real-time animation: it is not yet visually informative. But the population inevitably drops way below this number as the ecosystem stabilizes, leaving a smaller number of statistically better swimmers. After this, informative and enjoyable animations can be watched for hours.

2 Related Work