Robot shaping: developing autonomous agents through learning
Artificial Intelligence
Discontinuity in evolution: how different levels of organization imply preadaptation
Adaptive individuals in evolving populations
An Introduction to Genetic Algorithms
An Introduction to Genetic Algorithms
Unconstrained Evolution and Hard Consequences
Papers from an international workshop on Towards Evolvable Hardware, The Evolutionary Engineering Approach
Evolutionary Computation - Special issue on magnetic algorithms
Dynamic Subset Selection Based on a Fitness Case Topology
Evolutionary Computation
| Hi-index | 0.00 |
A number of authors have argued that redundancy in biological organisms contributes to their evolvability. We investigate this hypothesis via the experimental manipulation of genetic redundancy in evolving populations of simulated robots controlled by artificial neural networks. A genetic algorithm is used to simulate the evolution of robots with the ability to perform a previously studied task. Redundancy is measured using systematic lesioning. In our experiments, populations of robots with larger genotypes achieve systematically higher fitness than populations whose genotypes are smaller. It is shown that, in principle, robots with smaller genotypes have enough computational power to achieve optimal fitness. Populations with larger (redundant) genotypes appear, however, to be more evolvable and display significantly higher diversity. It is argued that this enhanced evolvability is a direct effect of genetic redundancy, which allows populations of redundant robots to explore neutral networks spanning large areas of genotype space. We conjecture that, where cost considerations allow, redundancy in functional or potentially functional components of the genome may make a valuable contribution to evolution in artificial and perhaps in biological systems. The methods described in the article provide a practical way of testing this hypothesis for the artificial case.