The following work is a companion piece to Karl Sims' paper, "Evolving 3D Morphology and Behavior by Competition".

With the intent to explore interesting spaces, Karl Sims developed a system in which he evolved competiting virtual creatures. In the past, the focus of evolutionary tactics were used in hopes of simple optimization. Sims' interest lied more in developing a diverse population that exhibited "interesting" behavior.

The actual competition consisted of two creatures, facing off in a physically simulated three-dimensional environment. The creatures were placed facing each other, equidistant from the center goal object, a cube. The object of the competition is for a creature to make its way closest to the cube and control as much of the cube as possible. This was done by at start time, activating the nervous sytem of the creature until a set time which determined the match had ended. The following equation was used to determine the fitness of a creature:

Fitness₁ = 1.0 + ((d₂ - d₁) / (d₁ + d₂))

Fitness₂ = 1.0 + ((d₁ - d₂) / (d₁ + d₂))

where d_i denotes creature i's distance from the cube.

Pictures of a competition starting [2]:

Sims hypothesized upon how different competition patterns would affect the evolution of creatures. In light of the fact that evolution of these creatures was performed on a Connection Machine (CM-5), a massive parallel processing supercomputer, processing time was still major contributing factor to how to setup each creature to compete. Sims describes four single-species patterns: all vs all, random pairing, tournament, and all vs best. All vs all takes every individual and plays them again all others in the population. However, this requires a large number of total competitions. Random pairing takes each individual and pairs them with a single opponent. While this pattern requires a small number of competitions, fitness is quite subjective since it is based on a single opponent. Sims describes this situation as, "... fitness can be more dependent on the luck of receiving a poor opponent than on an individuals actual ability." A tournament pattern allows each winning individual to compete against other winning individuals. Sims final suggestion, which he establishes as the more interesting pattern in his particular experiment, is for each individual to compete against the previously most fit individual in a "king of the hill" competition.

Sims' creatures were created using directed-graphs for a variable-length genotype. The creature develops by starting at the root node of the graph and following connections from this node to other nodes using the information contained therein. This allowed for cyclic and recursive parts. Sims goes on to describe how the information within the node allowed for various types of joints and parameters, along with child parts and their relative properties to their parents.

Not only was the physical structure of the creature evolved, but the underlying behavioral mechanism (the nervous system) was also evolved in conjunction. Sims brains of the creatures used input sensor values and produced output effector values. In between these inputs and outputs, neurons existed that could modify and provide calculations before being sent to various effectors. Sims used three sensors: joint-angle sensors which provided the degree of freedom of each joint, contact sensors that fired when a portion of the body of the creature was in contact with anything else, and photosensors which allowed the creature to "see" the cube and his opponent. Attached to these senors were neurons, which allowed computation to be performed on incoming senors. Finally, output was placed to effectors, which simply exerted a joint force.

Once a random population was generated, creatures are paired off, competitions are held, and the fitness of each creature is evaluated. From here on out, parents are selected based on fitness, children are bred through crossover, asexual reproduction, and mutation based on their parents genotypes. These possible children are then tested for survivability and survivors are kept for the following population.

Sims noticed in his results that different runs showed different evolutionary progress. Sometimes it took many generations for a creature to reach the cube at all, other times a successful strategy was discovered quickly. Aside from rate of evolutionary progress, interesting behavior developed amongst the creatures. Strategies for reaching the cube consisted of extending arms, falling on top of the cube, even crawling and rolling developed. When both competition creatures had evolved to reach the cube, strategies and counter-strategies developed as in swatting the cube away and chasing after it, smothering the cube to block access, and even offensive moves against opponents before snatching the cube. Below are some snapshots of competitions in play. The first involves an arm-like creature wrapping itself around the cube when facing a crab-like creature. The second shows a creature with arms trying to bat the cube to one arm.

Video Interview with Karl Sims - (-1:30min - Creatures from the Cube game)