Piet Van Remortel

Shrinking the Genotype: L-systems for EHW?

Inspired by biological development where from a single cell, a complex organism can evolve, we are interested in finding ways in which artificial development may be introduced to genetic algorithms so as to solve our genotype challenge. This challenge may be expressed in terms of shrinking the genotype. We need to move away from a oneto-one genotype-phenotype mapping so as to enable evolution to evolve large complex electronic circuits. We present a first case study where we have considered the mathematical formalism L-systems and applied their principles to the development of digital circuits. Initial results, based on extrinsic evolution, indicate that our representation based on L-systems provides an interesting methodology for further investigation. We also present our implementation platform for intrinsic evolution with development, enabling on-chip evaluation of grown solutions.

Lineage and Induction in the Development of Evolved Genotypes for Non-uniform 2D CAs

Biological development is a stunning mechanism that allows robust generation of complex structures from a linear building plan. This makes it an interesting source of inspiration for solving problems where direct manipulation of a higher-order structure is hard, and the generative building plan can be used as a substitute for indirect manipulation of the unfolded structure. In this paper we propose CA-DEV as a simple computational model for development of rules for non-uniform 2D cellular automata. While being a simplified version of more complex bio-inspired models, CA-DEV incorporates both lineage and induction, and is easily combined with artificial evolution through a binary genotype. We report an umber of basic experiments in evolving genotypes for CADEV with different settings related to cell division and induction. These experiments show that while the power to introduce diversity is high with most settings, structural properties of developed phenotypes are of adifferen t nature depending on the properties of the development adopted.

22 – Evolvable hardware: pumping life into dead silicon

The design productivity gap in the electronic industry is a well-known fact. How can the design community utilize the design capacity that technology is offering and at the same time ensure its correctness? To find solutions to the problem of developing large and complex designs new design paradigms are required (Semiconductor Industry Association, 1997). One possible solution is to turn away from traditional design techniques following the various design and testing phases and instead allow hardware to evolve until a correct solution is found. This technique is termed hardware evolution or equally, evolvable hardware. Evolvable hardware may be considered to be a subset of evolutionary computation, where the evolved solution is represented in hardware instead of software. In this chapter the field of evolvable hardware is introduced along with some of the limitations which prevent evolution of complex circuits. The process of natural development is analysed with respect to the features which may be represented in a developmental process for circuit development. The technology itself is also studied to identify features which development may exploit. We present some initial experiments for circuit development based on Lsystems, a mathematical formalism for development (Lindenmayer, 1968). The results of these experiments along with our analysis of development and our technology lead us to describe some of the challenges we face in our bid to develop electronic circuits.

Developmental effects on tuneable fitness landscapes

Due to the scalability issue in genetic algorithms there is a growing interest in adopting development as a genotype-phenotype mapping. This raises a number of questions related to the evolutionary and developmental properties of the genotypes in this context. This paper introduces the NK-development (NKd) class of tuneable fitness landscapes as a variant of NK landscapes. In a first part the assumptions and choices made in defining a simplified model of development genomes are discussed. In a second part we present results of the comparison of NK and two variants of NKd landscapes. The statistical properties of the landscapes are analysed, and the performance of a standard GA on the different landscapes is compared. The analysis is aimed at identifying the influence of the properties by which the landscapes differ. The results and their implications for the design of computational development models are discussed.