In machine learning and artificial evolution systems, the more interesting results, such as Sims' creatures or expert Backgammon players, are due more to elements of the learning environment than to any sophistication in the learning algorithm itself [132,108]. By keeping inductive biases and ad hoc ingredients to a minimum, we have demonstrated that interesting real-world behavior can come from a simple virtual model of physics and a basic adaptive algorithm.
The use of modular building elements with predictable -- within an error margin -- properties allows evolutionary algorithms to manipulate physical entities in simulation in ways similar to what we have seen, for example, in the case of robot control software. The bits in our artificial chromosomes are not limited to codifying just bits; they are capable of representing the building blocks of an entire physical structure.
We believe to have only scratched the surface of what is achievable. Combined with suitable simulators, the recombination of modular components guided by an artificial selection algorithm is a powerful framework capable of designing complete architectures ready to be built and used, discovering and exploiting complex properties of the substrate which are not identical to those explored by human engineers and designers.