J. C. Astor & C. Adami – A Developmental Model for the Evolution of Artificial Neural Networks
The model of the aligning hexagonal occupy-able “spaces” is an effective and understandable simplification. I like the attempt to model the many factors during neural development including simultaneous cell and connection (axon/dendrite) growth. A powerful feature is the inclusion of gene expression modification during the development/growth, producing cells that have the same genes but are producing different proteins and levels of each. This is a true dynamic simulation.
Also identifying low-level core network outcomes (self-limitation, logic gate development etc) is a powerful and useful analysis.
On the downside, the results section and conclusions are weak, I suspect because they have few because they have bitten off more than they could chew in terms of complexity. By incorporating so many “genes” and cross-dependent factors they have made the use of the clearly advantageous GA approach difficult because of the search space size. This is compounded by what appears to a belief that the grown network must be judged on fitness of performing more complex higher level behaviours. There is evidence that this growth stage simply sets the scene with a semi-blank canvas structure that can then adapt through plasticity to develop higher level and cognitive behaviours.
If they had included less factors they might have achieved similar results without having to resort to this complicated distributed GA architecture distraction. The justification for including so many factors up front is not clear – a more incremental approach might have been more effective.
Filed under: 1
