Zenith Grant Awardee

Christopher John Watkins

Royal Holloway, University of London

Co-Investigators

Susanne Still , University of Hawai'i at Mānoa; Lee Altenberg, University of Hawai'i at Mānoa

Project Title

Intelligence in Context

Project Summary

We will take a new look at two fundamental problems in evolution. First, the genome encodes both body-structure and – for small animals – much of their behaviour. Evidence from whole-genome studies shows that information about just about any particular aspect of our bodies is encoded at very many locations spread all over the genome. We term this a `diffuse’ genetic code. How do these diffuse genetic codes evolve, and how are they decoded? Nobody knows. We will work from first principles using information theory to construct examples of such diffuse codes that would be informationally efficient in evolution. Second, how does evolution combine with learning? For simple animals – think of beetles or small birds – their behaviour is intricate and adapted for their survival. Bur their behaviour seems mostly genetically determined with limited contributions from individual learning from experience. We will use a newly developed type of evolutionary simulation to construct evolutionary models in which individual learning and evolutionary learning are treated in the same mathematical way. We how this will enable us to understand the possible interactions of evolutionary and individual learning better; for example, we may be able to give principled general explanations of what determines how much of behaviour is learned by individuals and how much is encoded in the genome. We hope our work will be useful in gaining a general mathematical understanding of how our genomes determine complex anatomy, and of how innate knowledge interacts with individual learning.

Technical Abstract

We will consider the intelligence and development of simple organisms, with cognition far simpler than that of humans. We suggest that learning and intellectual development are tightly coupled to evolution, and that evolution, cognitive and physical development and behaviour should be understood together. We will combine three new theoretical tools to gain a deeper understanding of the relationship of learning and evolution: a class of genetic algorithms that satisfy detailed balance, and which are closely related to Bayesian nonparametric MCMC; an information theoretic measure of precision of adaptation, which we will use to assess the informational efficiency of genetic encodings of complex traits; and thermodynamic analysis of dissipation and predictive inference. We will develop evolutionary systems in which evolutionary and individual learning are part of the same theoretically tractable probability model. We will develop a new test-bed of evolving 'digital organisms' that are efficiently implemented as probabilistic finite state machines, in which to demonstrate evolution of metabolism and simple agency.