Molecular self-organisation in a developmental model for the evolution of large-scale artificial neural networks
Hamid Bolouri, Rod Adams, Stella George, Alistair G. Rust · University of Hertfordshire Research Archive (University of Hertfordshire) · 1998
We argue that molecular self-organisation during embryonic development allows evolution to perform highly nonlinear combinatorial optimisation. A structured approach to architectural optimisation of large-scale Artificial Neural Networks using this principle is presented. We also present simulation results demonstrating the evolution of an edge detecting retina using the proposed methodology. Introduction One of the attractions of Artificial Neural Networks (ANNs) has been the possibility of designing intelligent systems capable of optimising their functionality according to application requirements. Adapting the architecture of an ANN (number of neurons, connectivity pattern, and neuron function) to a given application can be viewed as a combinatorial optimisation problem. For sophisticated applications, the problem tends to be high-dimensional and highly nonlinear. Direct applications of current combinatorial optimisation methods to such problems tend to be unacceptably inefficient...