Master numerical tasks with ease

AT THE END OF EVERY semester, student grades are due 72 hours after the final exam. When grading is done, teaching assistants put together the final spreadsheet with cumulative scores, computed as weighted averages of scores for homework, course projects, and exams. My task is to assign letter grades to individual students an attempt at fairness while keeping the course attractive to future generations and yet avoiding grade inflation. When assigning grades ‘‘on the curve,’’ it is common to start with historical quotas for A’s, B’s, and C’s for the course the mean can be anywhere from B or Bþ, and a quarter of students usually receive grades from A to Aþ. Students with exceptionally low scores are given exceptionally low grades, and specific percentile boundaries for common grades are adjusted based on the actual workload in this semester. To implement such a policy, the rows of the spreadsheet are sorted by cumulative scores so that percentile boundaries can be translated into absolute boundaries. Regrettably, absolute grade boundaries can look arbitrary, and a careless mistake on a simple homework question can, in principle, turn an A into a Bþ in a student’s transcript. To make the letter grades more stable, the boundaries are further relocated up or down to the closest gap between successive scores (the gaps are computed and recorded next to scores). As a side effect, this step allows us to be generous with re-grades on exam questions small changes in scores rarely affect letter grades. Before entering grades, we count A’s, B’s, and C’s to compare them against the original quotas, and perform a few spot checks. EDA-minded readers may raise several interesting questions. To what degree can the grade-assignment process be automated? How should this automation be implemented and maintained? How can results be verified? Which tools or environments should be used to visualize results, for example, in the form of histograms? What pricing model is most appropriate for such automation? Let us examine the traditional approach perfected by the EDA industry a new stand-alone software product, which we would code-name GradeCompiler. GradeCompiler would read a dedicated input file format with separate sections for raw scores, overall grade quotas and a course-complexity adjustment. Grade-boundary determination would be formalized as an NP-hard optimization problem and solved by fast heuristics. Individual student grades could then be listed in an output file. A prototype implementation written in Cþþ by two graduate students could be commercialized, extended with a GUI and visual histograms, and maintained on several OS platforms. A separate verification tool would ensure that actual grade statistics closely match those in the input file. Given the potential market of over 400,000 educators, a sales force would need to be assembled. License management software would limit software piracy, and support contracts would improve adoption while increasing the return on investment. Pricing could be based on recurring usage (number of courses) or end result (number of students with grades) one model may be appropriate for smaller colleges, and the other for large, established universities. Given a double-digit growth in US tuitions, universities could be expected to afford Reviewed in this issue Advanced Excel for Scientific Data Analysis, 2nd ed., Robert de Levie (Oxford University Press, 2008, ISBN: 978-0-195-37022-5, 736 pp., $59.50).