The Artificial Ontogeny Project

Project Description

This project is one of the Embodied Artificial Evolution research projects. This project deals with isolating principles from biological development that are useful for evolving modular agents in a physically realistic simulation. Modular agents refer to agents, or robots, that are constructed from a series of similar or identical units, possibly with corresponding modular neural structure. One aspect of this project is to elucidate the relationship between a modular genetic regulatory system, in which gene duplication and differentiation leads to dynamic coupling and dissociation of genetic traits, and a modular phenotype.

The goal of this work is to use the lessons learned to prove that the resultant developmental encoding scheme exhibits three characteristics which explain its power for evolving modular agents. First is that the encoding scheme is highly evolvable, meaning that artificial evolution continually explores the design space; it is compact, meaning that convergence to a useful solution is rapid; and that it is robust, meaning that the evolved agents can perform well in different environments, including a transformation from a simulated agent to a real-world robot.

A simulation system, called Artificial Ontogeny, has been written, which combines artificial evolution with artificial development. A variable-length, floating point genetic algorithm has been integrated with a three-dimensional, physically realistic simulation package. The genomes of the genetic algorithm encode genetic regulatory networks, which direct the growth of a virtual agent in the simulation package. The agent is then tested against a given fitness function. The simulation is described in (Bongard & Pfeifer, 2001). In (Bongard & Pfeifer, 2001) we demonstrated that the developmental encoding is compact, and how such an encoding leads to repeated, higher-order phenotypic structures in the evolved agents. In a forthcoming publication we will show in more detail how evolution shapes the underlying genetic regulatory networks of the virtual agents. Finally because the encoding scheme is highly evolvable, it will be shown that it can be used to evolve agents which exhibit increasingly non-trivial behaviour, which an outside observer may classify as intelligent behaviour. Thus this work may contribute to embodied cognitive science by providing examples of the evolution of intelligence in a virtual environment.

Fig. 1: A small, locomoting creature	Fig. 2: A large agent with repeated structure

Links

Downloads   Interactive demonstrations of evolved agents.
Karl Sims    The first researcher to evolve both the morphology and neural control of complete agents.
The Golem Project    Transferring evolved robots from simulation into the real world.

References

Bongard, J. C. & R. Pfeifer (2001), Repeated Structure and Dissociation of Genotypic and Phenotypic Complexity in Artificial Ontogeny in Spector, L. et al (eds.), Proceedings of The Genetic and Evolutionary Computation Conference, GECCO-2001. San Francisco, CA: Morgan Kaufmann publishers, pp. 829-836.

Bongard, J. C., & Paul, C. (2000), Investigating Morphological Symmetry and Locomotive Efficiency using Virtual Embodied Evolution in Meyer, J.-A. et al (eds.), Proceedings of the Sixth International Conference on Simulation of Adaptive Behaviour, pp. 420-429.