Artificial Life research brings together methods from Artificial Intelligence (AI), philosophy and biology, studying the problem of evolution of complexity from what we might call a constructive point of view, trying to replicate adaptive phenomena using computers and robots.
Here we wish to shed new light on the issue by showing how computer-simulated evolutionary learning methods are capable of discovering complex emergent properties in complex domains. Our stance is that in AI the most interesting results come from the interaction between learning algorithms and real domains, leading to discovery of emergent properties, rather than from the algorithms themselves.
The theory of natural selection postulates that generate-test-regenerate dynamics, exemplified by life on earth, when coupled with the kinds of environments found in the natural world, have lead to the appearance of complex forms. But artificial evolution methods, based on this hypothesis, have only begun to be put in contact with real-world environments.
In the present thesis we explore two aspects of real-world environments as they interact with an evolutionary algorithm. In our first experimental domain (chapter 2) we show how structures can be evolved under gravitational and geometrical constraints, employing simulated physics. Structures evolve that exploit features of the interaction between brick-based structures and the physics of gravitational forces.
In a second experimental domain (chapter 3) we study how a virtual world gives rise to co-adaptation between human and agent species. In this case we look at the competitive interaction between two adaptive species. The purely reactive nature of artificial agents in this domain implies that the high level features observed cannot be explicit in the genotype but rather, they emerge from the interaction between genetic information and a changing domain.
Emergent properties, not obvious from the lower level description, amount to what we humans call complexity, but the idea stands on concepts which resist formalization -- such as difficulty or complicatedness. We show how simulated evolution, exploring reality, finds features of this kind which are preserved by selection, leading to complex forms and behaviors. But it does so without creating new levels of abstraction -- thus the question of evolution of modularity remains open.