We have described in detail some of the solutions encountered by our test domains: the appearance of nested levels of complexity in Lego structures, the reuse and adaptation of components, the emergence of basic navigation behaviors such as wall-following, spiraling, maze navigation and space filling, amongst Tron agents.
The Brooksian idea that an agent's complexity sometimes comes from the environment more than the agent itself is supported by the evidence of Tron robots: these stateless agents have no internal means for storing plans or previous decisions, yet were shown (section 3.8) to reliably evolve behaviors involving choreographed sequences of moves.
By adapting to the real problem, rather than a metaphor or an incomplete simulation, the applied aspect of evolutionary methods comes to life, which could lead to a variety of applications.
Fully Automated Design, the idea of computers designing complete, functional artifacts, based on constraints and goals defined externally, was shown here with Lego structures. Lipson and Pollack extended this idea by showing how a static motion simulator can be used to coevolve morphology and brain of walking creatures . Automated design is central to evolutionary robotics, for brains need bodies to inhabit, as well as evolutionary design, which needs to break the mold of pre-imposed problem decomposition, and play more freely with components and goals.
Together with work by other researchers such as Sims and Thompson our thesis deals with what we have called the reality effect: when evolution interacts with a large, complex environment like those typically generated by the world around us, complex solutions appear that exploit emergent properties of the domain in surprising new ways.