H.H. Bauer's knowledge-filter model of scientific progress makes clear where I think that science can be improved with the use of computational discovery programs: not so much in the application of textbook knowledge, but rather at the science frontiers where contributions get made to the primary (journal) literature. This literature is filled with often unreliable inferences from experimental evidence to models, patterns, empirical descriptions, and the like. At this stage, both knowledge-driven and data-driven programs can make significant contributions to making a variety of inferences in frontier science more reliable, as well as faster and maybe even cheaper.

This filter model resolves the seeming paradox: whereas much science is highly reliable (it puts Man on the moon, it cures many diseases), this body of knowledge is textbook science which largely consists of knowledge of the form "if we do X, Y will happen." Frontier science - tomorrow's textbook science - may put Man on Mars, but only after much filtering because of its inherent unreliability. This is where a scientist/computer collaboration can accelerate progress in frontier science.

Reference: Scientific Literacy and the Myth of the Scientific Method, Henry H. Bauer, University of Illinois Press , Urbana IL, 1994.