ALife is founded upon a different paradigm than classic AI: it considers human intelligence as a small part logic, and a large part interaction with the habitat -- embodiment. Evolutionary methods have shown the potential to develop artificial entities adapted to complex habitats and discover their rules. This is, however, only part of the problem, because it is not known how to abstract those found rules and incorporate them into the agent as modules, as new words in the language.
Here we have shown that recombination plays a fundamental role, acting as a simple way of finding and reusing higher-level descriptions and subsolutions. One of the reasons why agents can find and exploit emergent rules on their domains is that there is a partial reduction of complexity coming from the crossover operator, which replicates useful components at random. The question for the future is: how can these emergent rules be assimilated by the agent? ALife has only started to look at this problem, sometimes referred to as the modularity problem [5,74].
The bulk of evidence coming from ALife work, including the one described here, supports the g-t-r hypothesis of Universal Darwinism: it is plausible that selection and recombination, in the scope of a challenging environment, can lead to increasing levels of complexity. But it also shows that this is still a crude model of evolution. Two big questions remain:
Take the genome of mammals as an example. It is organized to protect a core part of the genotype, while encouraging mutation on regulation factors. The result is a basic template that allows the exploration of morphology without changing the basic components . Horizontal gene transfer, which leads to inter-species recombinations, has been controlled by sexual reproduction, to make it happen only within the species. Thus not only emergent complex components have been found, such as organs, but they have been incorporated into the representation, and the language of mutation and recombination itself has been reformulated.
Coevolution (as a subfield of evolutionary computation) has obtained interesting results that could be pointing in the right general direction. These experiments do not have a drastic agent/environment distinction. Instead, the environment is the result of the collective behavior of the population. This configuration, when successful, creates a continuous challenge at the right level, leading to agents that climb the ladder of complexity. Coevolution thus breaks the first mold, that of a changing agent in a rigid environment. But real evolution involves simultaneous co-adaptation throughout all levels of organization, from the individual gene to the symbiotic association of species.