Summary: Contributions of this Thesis

Here we investigate the reality effect, of which previous works in the field (Sims' virtual creatures, Thompson's FPGA's) are examples: evolution interacting with reality discovers emergent properties of its domain, builds complexity and creates original solutions, resembling what happens in natural evolution.

We investigate two complementary scenarios (a simulation that brings a computer brain out into the real world, and a video game which brings a multitude of natural brains into a virtual world) with two experimental environments:

1.

Evolution of structures made of toy bricks (chapter 2). These are the main points discussed:

• Evolutionary morphology is a promising new domain for ALife.
• Adaptive designs can be evolved that are buildable and behave as predicted.
• Principles of architectural and natural design such as cantilevering, counterbalancing, branching and symmetry are (re)discovered by evolution.
• Recombination between partial solutions and change of use (exaptation) are mechanisms that create novelty and lead to the emergence of hierarchical levels of organization.
• Originality results from an artificial design method that is not based upon pre-defined rules for task decomposition.
• The potential for expanding human expertise is shown with an application -- EvoCAD, a system where human and computer have active, creative roles.

2.

Evolution of artificial players for a video-game (chapter 3). Main issues are:

• Evolution against live humans can be done with a hybrid evolutionary scheme that combines agent-agent games with human-agent games.
• The Internet has the potential of creating niches for mixed agent/human interactions that host phenomena of mutual adaptation.
• Basic as well as complex navigation behaviors are developed as survival strategies.
• Coevolving in the real world is stronger than coevolving in an agent-only domain, which in turn is stronger than evolving against a fixed training set.
• Statistical methods are employed in order to analyze the results.
• Agents adapted to complex environments can exhibit elaborate behaviors using a simple reactive architecture.
• Human learning arises and can be studied from the interactions with an adaptive agent.
• An evolving population acts as one emergent intelligence, in an automated version of a mixture of experts architecture.

We conclude with a discussion on AI and the role of discovery and of interaction between learning algorithms, people and physical reality in the light of these results (chapter 4).

Altogether, we are elaborating on a new perspective of Artificial Life, conceived as one of the pieces on the question of evolution of complexity. The evolutionary paradigm explains complexification up to a certain point at least, but also shows that we are still far from a complete understanding of this phenomenon.