The goal of this thesis is to show how the dynamics of computer-simulated evolution can lead to the emergence of complex properties, when combined with a suitable environment. We propose that one way to do this is to put evolutionary algorithms in contact with the real world, precisely because it was in this context that natural evolution led to the sophisticated entities conforming the biosphere.
Even though the characteristics that define ``suitable environment'' for evolution are unknown, we should be able to verify the theoretical predictions of the evolutionary hypothesis by placing artificial agents in the same kinds of contexts that produce complex natural agents.
The difficulty of measuring complexity makes it hard to study an evolutionary system acting on a purely symbolic domain, such as the Tierra experiments [112,113]. Evolving real-world agents instead makes it easier to recognize solutions to difficult problems which are familiar to us, and at the same time creates an applied discipline, dealing with real problems.
We are deliberately staying away from a discussion about the different flavors of evolutionary algorithms (Genetic Algorithms, Genetic Programming, Multi-objective optimization and so on): all of them capture the fundamental ideas of the g-t-r model. Our aim is to reproduce the dynamics of natural evolution of complexity by situating artificial evolution within complex, reality-based domains. We are driven by the intuition that the most interesting results in our field have come not from great sophistication in the algorithm, but rather from the dynamics between g-t-r and interesting environments. Exciting results using coevolution [65,131,109,123] for example, suggest that the landscape created by another adaptive unit is richer than a fixed fitness function.
Previous work in Artificial Life has already shown promising examples of evolving in real worlds. Here we implement two new approaches: the first one is evolving morphology under simulated physical laws, with simulated elements that are compliant to those found in reality, so as to make the results buildable. The second approach is to employ the concept of virtual reality to bring living animals into contact with simulated agents, in order to evolve situated agents whose domain, albeit simulated, contains natural life forms.