Canalization captures both cell differentiation processes during development as well as the reproducibility of traits despite environmental, developmental or genetic noise observed in wild-type organisms that have been under natural selection. On the one hand, canalization tends to keep phenotypes reproducible, but on the other hand, it is evident that phenotypic evolution has taken place in the past. We need to understand the role of canalization in an evolutionary context, specifically, how canalized traits could evolve. In this thesis, I address three questions that are related to different aspects of this question.,First, considering highly canalized traits at the molecular level, to what degree do variations exist that could pass through the sieve of development? What is the extent of phenotypic variation within a species? This question has been little studied because the conservation of the expression of these patterning genes across species gave rise to the idea that little variation could be expected within a species. In Chapter 2, I consider the the expression of an essential developmental patterning gene, even-skipped. I address this question by observing the expression of this gene in inbred lines derived from natural populations. The results demonstrate that appreciable expression variation exists within natural populations. In addition, I was able to characterize the molecular cause for one particular expression variant.,Second, given a trait that is highly conserved across species, are the same or different mechanisms being used to generate the same end phenotype? For a canalized trait which is conserved across large evolutionary distance, in this case even-skipped expression in Drosophila melanogaster (D. mel) and the sepsid fly Themira putris (T. put), there is good evidence that expression is conserved in the face of significant divergence of regulatory sequence. We address this question using a computational model of transcription to help to understand the possible mechanism that drives different enhancer expression from two different species. Theoretical analysis indicated that Caudal binding sites had replaced those for Bicoid in the sepsid enhancer driving stripe 2. In Chapter 3, I test this conclusion by constructing embryos that had no Caudal protein and performing a comparative assay of the expression of even-skipped stripe 2 driven by enhancers from two species. I found the T. put stripe 2 enhancer reduces its expression much more than the D. mel stripe 2 enhancer in the cad background, supporting the computational prediction.,Finally, in Chapter 4, I address the general question of how canalization could evolve using a computational model. I do this by simplifying canalization to a one-step process of genotype to phenotype map, but capturing its important property---robustness by a parameter in the model which can be explicitly controlled, either by the experimenter or by a genetic locus that can itself evolve. Fitness is determined by both individual's phenotype and the environment. I found high robustness has the selection advantage over low robustness populations under a stable environment to which the organism is optimally adapted, while low robustness has the selection advantage over high robustness ones when adapting to a new environment. The phenotypic space for high robustness populations is smaller than low robustness populations. When robustness itself is controlled by a genetic locus such that it evolves with the phenotype, low robustness is selectively advantageous in a transient manner immediately after an environmental shift. Shortly thereafter, however, and long before the population is well adapted, high robustness becomes selectively advantageous.