Bridging the Gap: Teaching Computational Intelligence to the Next Generation of Engineers

TL;DR

This article explores the transformation of COMP307, an undergraduate course at Victoria University of Wellington, from a classical symbolic AI elective into a comprehensive, high-demand program. By balancing traditional logic with Neural Networks and Evolutionary Computation, the course provides students with the "Neural Engineering" skill set required to solve real-world industrial optimization and classification problems.

Problem & Motivation: The AI Education Bottleneck

For years, undergraduate AI education remained siloed in the realm of symbolic logic (Prolog, search algorithms, and rule-based systems). While these are fundamental, the modern engineering landscape demands a shift toward Computational Intelligence (CI).

The challenge faced by the authors was twofold:

1. Resource Scarcity: Lack of curriculum space to offer separate courses for Machine Learning, Neural Networks, and Evolutionary Computation.
2. Practicality Gap: Students often learn the math of CI but struggle to apply it to "messy" real-world data.

The authors hypothesized that a consolidated, hands-on course could serve as a "platform" for both aspiring researchers and future industry professionals.

Methodology: A Dual-Core Curriculum

The course structure is a carefully orchestrated 12-week journey split into two halves: Symbolic Intelligence and Machine Learning/CI.

The CI Component Breakdown:

• Simple Learning: Grounding students in the foundations (k-NN, Naive Bayes, Decision Trees).
• Neural Engineering: Moving beyond the Perceptron to Multi-Layer Feed-Forward Networks. The focus is not just on derivation but on engineering choices—architecture selection, hyperparameter tuning, and preventing over-fitting.
• Evolutionary Computation (EC): Introducing Genetic Algorithms (GA) and Genetic Programming (GP) as powerful optimization tools, contrasting their search heuristics with gradient-based methods.

Figure 1: The course serves as a bridge for students to apply CI methods to industrial applications.

Experiments & Results: The Power of Implementation

The core of the paper’s success metric lies in its Individual Assignments. Unlike courses that rely on black-box libraries early on, students are required to:

1. Code from Scratch: Implement (k-)NN, Naive Bayes, and Perceptrons in languages like Java or Python.
2. Benchmark Testing: Use the UCI Machine Learning Repository to test their code against tasks like intrusion detection and medical diagnosis.
3. Neural vs. Evolutionary Analysis: Compare the complexity of evolved Genetic Programs against Decision Tree rules.

Key Outcomes:

• Scale: COMP307 grew to be the largest 300-level course at the university since 2009.
• Student Engagement: Despite a heavy workload (150 hours total), engineering students reported that the CI techniques were the most "useful" modules for their specific disciplines.
• Research Pipeline: The course successfully feeds students into Honors and PhD research, specifically in Evolutionary Computation.

Critical Analysis & Conclusion

The experience at Victoria University demonstrates that breadth does not have to sacrifice depth at the undergraduate level. By focusing on the intuition of "Neural Engineering"—treating the network as a system to be designed and tuned—students gain a professional-grade capability.

Takeaway: The success of the COMP307 model lies in its Inductive Bias toward practical application. Rather than just teaching "What" an algorithm is, it teaches "How" to make it work in the face of real-world noise.

Limitations: As the paper notes, as an undergraduate course, it cannot cover the absolute SOTA (State-of-the-Art) in deep learning (which would require a 400-level specialization). However, it provides the essential "mathematical and algorithmic scaffolding" required to reach that level.

Future Outlook: For educators, the next frontier will be integrating Generative AI and Transformer architectures into this existing framework without losing the fundamental lessons of the "Simpler" methods that COMP307 so effectively teaches.