120 likes | 135 Views
A System to Place Incoming Students in Classes. {. Henry M. Walker, Andrew Hirakawa, Russel Steinbach Grinnell College, Grinnell, Iowa. 1. Three specific objectives: Identify most appropriate course Communicate placement to students, advisors Publicize interesting courses, recruit.
E N D
A System to Place Incoming Students in Classes { • Henry M. Walker, Andrew Hirakawa, Russel Steinbach • Grinnell College, Grinnell, Iowa 1
Three specific objectives: • Identify most appropriate course • Communicate placement to students, advisors • Publicize interesting courses, recruit Problem: Place incoming students in CS, math, statistics 2
Based on SAT, ACT, similar aptitude tests • Often required for admission • Tests measure aptitude, not background • Placement tests • Numerous tests commercially available • Logistical issues in administration • calibration needed with local courses • Placement by faculty • Can tap faculty expertise • Time consuming • Variation possible from different faculty • Use of an expert system • Available for incoming, prospective students • Allows follow-up studies, based on performance • Needs non-trivial development time Placement Approaches 3
Use available student data • Standardized tests: SAT, ACT, AP, IB • High school transcript • Base placements on rules • Initial rules from faculty experience • Follow-up studies: courses taken, grades • Inference engine applies rules to data • Develop placement letters, Web pages • Develop software • Version 1 (1993) LISP-based, TMYCIN • Version 2 (2010) PHP, locally developed Basic Approach for an Expert System 4
1993 version (LISP based) • TMYCIN Inference Engine: Gordon Novak, AI Lab, University of Texas at Austin • Vikram Subramaniam, Ivan Sykes • Consultants • 1993-2004: Eugene Herman, Charles Jepsen • 2004-2009: Emily Moore • 2009-present: Shonda Kuiper, Chris French, Karen Shuman • 2010: Barbara Johnson • Current version • Grinnell CS faculty • Dean’s Office for summer “MAP” funding Acknowledgments 5
Our system operates in two primary modes: • Generation of letters to students based on data from the Registrar. • Sample Student LaTeX Letters • 2. Web-based, tentative placements to • prospective students. • Web Placement Interface Outputs & System Overview 6
Data for incoming students • Files from Registrar’s Office • Storage in MySQL database • Web-based materials • Data from HTML forms (not stored in MySQL) • Programming in PHP • Connection with MySQL • Web scripting • Associative arrays for rule storage • Inference engine • TMYCIN (1993) ==> PHP (2010) • Backtracking algorithm for rule processing • Notes • Formal grammar for rule structure • Rules in LISP/PHP format (for scripting) • Rules printable in English format (faculty) Technical Overview 7
Rules composed have three parts • Name • Conclusion • Conditions • Examples: • array("statRule120","TSTAT-PLACE = 208", • array("all","stdscores = unknown", • array("some","statsem >= 2", • "apstat >= 1"))); • array("satRealRule10", "satmathreal =field satmath", • array("all", "satmath >= sat2math1", • "sat2math2adj < satmath")); Rule structure 8
Each Condition Array has two parts • Quantifier (e.g., some, all) • Conditions to be satisfied • Simple conditions • array("satRealRule10", "satmathreal =field satmath", • array("all", "satmath >= sat2math1", • "sat2math2adj < satmath")); • Condition array • array("statRule120","TSTAT-PLACE = 208", • array("all","stdscores = unknown", • array("some","statsem >= 2", • "apstat >= 1"))); Rule Conditions 9
Two types of conclusion • Fixed value conclusions • "stdscores = unknown" • Variable values • "satmathreal =field satmath" • Three parts • Field • Operator • Value Rule Conclusions 10
Short circuit evaluation • Rules are marked • Minimize SQL queries. Efficiency 11
Future Work: • Ever changing curricula both high school and college demand changes in inference rules. • Occasional re-evaluation of student performance versus generated placements. • Conclusion • Automated placements are a viable alternative. • Two experiments support this claim: • One experiment showed individual placements of students by faculty as reliable as nation-wide placement test. • Second experiment showed rule-based system placements were comparable with placements by faculty. Conclusion & Future Work 12