Communications of the ACM
Learning the Fourier spectrum of probabilistic lists and trees
SODA '91 Proceedings of the second annual ACM-SIAM symposium on Discrete algorithms
Efficient noise-tolerant learning from statistical queries
Journal of the ACM (JACM)
Machine Learning
Journal of the ACM (JACM)
A Complete Characterization of Statistical Query Learning with Applications to Evolvability
FOCS '09 Proceedings of the 2009 50th Annual IEEE Symposium on Foundations of Computer Science
Partial observability and learnability
Artificial Intelligence
Distribution free evolvability of polynomial functions over all convex loss functions
Proceedings of the 3rd Innovations in Theoretical Computer Science Conference
IJCAI'11 Proceedings of the Twenty-Second international joint conference on Artificial Intelligence - Volume Volume Two
MFCS'07 Proceedings of the 32nd international conference on Mathematical Foundations of Computer Science
Distribution free evolvability of polynomial functions over all convex loss functions
Proceedings of the 3rd Innovations in Theoretical Computer Science Conference
Hi-index | 5.23 |
The evolvability framework is a computational theory proposed by Valiant as a quantitative tool for the study of evolution. We explore in this work a natural generalization of Valiant's framework: an organism's genome is regarded as representing the Fourier spectrum of a real-valued function that the organism computes. A performance function is suggested that averages in a certain way the organism's responses over the distribution of its experiences. We show that this generalization supports the existence of an efficient, conceptually simple and direct evolutionary mechanism. More concretely, we consider the case where the ideal behavior that an organism strives to approximate is encoded by some decision list, and establish the evolvability of decision lists with respect to the suggested performance metric, over the uniform probability distribution. In accordance with biological evidence on how genomes mutate, the evolutionary mechanism we propose performs only simple operations on the organism's genome to obtain mutated genomes. The surviving genome is selected greedily among genomes in the current generation based only on performance. A sustained performance improvement is ensured, at a fixed and predictable rate across generations, and a highly fit genome is evolved in a number of generations independent of the size of the ideal function, and determined only by the required approximation degree. Furthermore, the size of the genome grows logarithmically in the number of environmental attributes. None of these rather stringent, and presumably biologically desirable properties are enforced by the baseline evolvability framework, nor are these properties possessed by other early evolvability results.