An early AI advising prototype shows why similar students still need different advice

A historical experiment, newly digitised

Choosing university electives involves more than finding an interesting subject. Students need the right prerequisites, a manageable workload and courses that fit their goals. Academic advisors help reconcile these demands, often while answering many routine questions.

Tanat Tonguthaisri explored whether software could assist with this work in an August 2003 Master of Engineering thesis at RMIT University, Enhancing Academic Advisement with Case-Based Reasoning. The arXiv posting dates from November 2026, but it presents a digital edition of that historical research—not new experiments or a fresh evaluation of the software.

The prototype addressed a narrower question than comprehensive academic counselling: could stored examples help recommend electives in a simplified engineering program?

Advice by comparison

Case-based reasoning means tackling a new problem by retrieving similar examples and reusing or adapting their solutions. Here, a case paired a student profile with a set of suggested courses.

The prototype represented students through three main attributes: their ranked interests across five technical fields, an overall academic-performance level and a preference for broad knowledge versus specialisation. These numerical descriptions let the software compare a new profile with stored cases.

This approach was not a conversational language model. It was a PHP and MySQL prototype that searched structured records and calculated similarity scores.

Three matching methods were compared. One measured how closely numerical profiles pointed in the same direction; another measured their distance apart. A combined method treated subject interests separately from performance and specialisation preference. The underlying question was practical: which way of defining “similar” produced more sensible course suggestions?

Similarity was only part of the decision

The system retrieved its ten highest-ranked cases and assembled recommendations from their course lists. Courses inherited scores from the strongest matching cases in which they appeared, with weighted occurrence counts intended to help break ties.

Rules then mattered. A subject with unmet prerequisites was marked unavailable. Another could be discouraged because its assigned difficulty exceeded the student’s performance category. The interface distinguished recommended courses, alternatives and these restrictions, with short explanations.

That division is central to the thesis: explicit rules handle requirements, while comparisons between cases help with preferences. A strong resemblance to another student does not make a missing prerequisite disappear.

Nor was a displayed matching percentage a probability of passing. It described the system’s numerical comparison. The thesis itself cautioned against comparing raw matching percentages across its different algorithms.

What the validation actually showed

The validation used specified test profiles and expected recommendations described in the thesis. It included eight scenarios testing difficulty restrictions, eight testing unmet prerequisites and two examples combining those constraints. A separate check examined whether removing some stored interest profiles preserved the relative ordering of the remaining cases.

The author reported that the combined matching method performed better overall on the comparison measures used. Some recommendations matched the expected course sets; others differed in membership or order.

One combined-constraint example makes the limitation concrete. A student interested in microwave engineering, networking and satellites could not take several relevant subjects because of difficulty or missing prerequisites. The prototype then recommended optics and signal-processing options ahead of an available satellite subject that the author considered more appropriate.

The thesis attributed such discrepancies partly to sparse case coverage and suggested additional rules. That explanation identifies a plausible weakness, but does not establish that expanding the library would reliably fix it.

These were prototype plausibility checks, not evidence from a deployment measuring graduation rates, student satisfaction or advisor workload. They also did not establish superiority over human advisors through an independently assessed trial.

Ambition beyond the evidence

The design used a reduced course structure, concentrated on a single stream and initially represented students progressing without failed courses. Its broad performance category could not capture every strength, weakness or exceptional circumstance.

The thesis envisaged wider availability and less repetitive work for advisors, even suggesting that a mature system might eventually replace them. Those were aspirations, not demonstrated outcomes. Its own limitations section advised crosschecking recommendations with an academic advisor and reserved important or unusual decisions for human attention.

What the edition can—and cannot—verify

The 2026 edition discloses AI assistance with transcription, proofreading, formatting and figure reconstruction. It reports page-by-page comparison with source images, while explicitly stating that this does not validate historical claims, calculations, citations or whether the PHP listing runs. One replacement diagram is a 2026 redraw, not a facsimile of the original.

This account relies on extracted text, which does not preserve every visual detail of figures and wide tables. Separate correction and unresolved-error records are mentioned but are not included in the supplied evidence. The defensible conclusion remains modest: the thesis demonstrates a possible combination of example-based recommendations and explicit constraints, while leaving real-world effectiveness unproven.