Evolutionary Robotic and Open-Ended Design Automation Lipson 2004

This paper is fantastic. Important ideas:

the use of L-systems is challenging. resuse of parts as the roots of structural modularity and reuse.

I've been inclined to wonder how to make GAs as effective at searching complex spaces as nature, but it could be that GAs are just as complex as evolution, andour ineffectiveness comes from inadequate narnessing of the same design principles. In other words, Nature's GA's don't stand out for being able to search relatively higher dimensional spaces, but for choosing the right design units.

Is the genotype phenotype distinction just a distraction for GA researchers or does respecting it really lead to better results?

I like the distinction and then tradeoff between modularity and regularity (or reuse and repetition, or specialization and repetition, or modularity and optimality). The tradeoff is through the notion of coupling.

Look up 'long jumps' and global variables as they relate to the optimality/modularity tradeoff.

"The increated performance gained by reduction of modularity is often justified in the short term, whereas increased modularity is often jusfified in longer time scales where adaptation becomes a dominant condieration" How to select for modularity. That is my research question.

Long quote: "It is not clear whether modularity, regularity and hierarchy are properties of the system being evolved (i.e. the 'solution'), or of the target fitness specification (i.e. the 'problem'). It may well be that there is a duality between these viewpoints. The evolutionary computation literature contains several instances of test functions that are temselves modular (separable, e.g. Royal Roads), hierarchical (e.g. Hierarchical-IFF), and reular (e.g. one-max). It is not surprising then to see corresponding algorithms that are able to exploit these properties and find the solutions to these problems. " This is rough. Again, how to select for the structures (and what does it mean to do so?).

".. It is therefore plausible that in search of scalable algorithms for synthesizing solutions bottom up, we should aboid test functions that have an inherent modular or hierarchical rewared, and have these solution preoperties emerge from the search process itself"

Good position quote: "It is clear that the complexity of engineering products is increasing to the point where traditional design precesses are reaching their limits. More manpower is being invested in managing and maintaining large system than designing them, and this ratio is likely to increase because no single person can fathom the complexities involed."

To Read:

• hartwell Hopfield Leibler and murray 1999 "From molecular to modular cell biology" nature
• Complex Adaptations and the evoluation of evolvability" Wagner, Altenberg 1996
• Finding building blocks through eigenstructure adaptation" Wyatt Lipson 20032004