5.4 Modeling and Simulation

Modeling and simulation in learning engineering encompass two related but distinct practices. Modeling refers to the construction of formal representations of learner knowledge states, cognitive processes, learning sequences, or system architectures — representations that make implicit assumptions explicit and enable prediction, testing, and refinement. Simulation extends modeling into dynamic, interactive contexts where learner-system interactions can be evaluated before or alongside real-world deployment. Together these practices provide learning engineers with tools for systematically testing design hypotheses, optimizing adaptive algorithms, and validating learning experiences against desired outcomes without requiring every hypothesis to be tested exclusively through costly large-scale deployments. [LE-LS-GL-007]

Student and Knowledge Modeling

The most consequential modeling tradition in learning engineering is student modeling — formal probabilistic representations of what individual learners know and can do, inferred from their observable behavior. Bayesian Knowledge Tracing, developed by Corbett and Anderson, is the field's foundational student model: it estimates the probability that a learner has acquired a given knowledge component based on the pattern of correct and incorrect responses across practice items, updating its estimate with each new response. [LE-LS-AP-011] This model makes the learner's latent knowledge state — not directly observable — tractable for algorithmic decision-making, enabling adaptive systems to present appropriate problems, trigger hints at the right moments, and determine when mastery criteria have been met. The full architecture of Cognitive Tutors, documented by Corbett, Koedinger, and Anderson, integrated student modeling with domain modeling (production rules encoding expert knowledge) and pedagogical modeling (strategies for selecting problems and providing feedback) into a coherent system. [LE-LS-AP-002]

The SOAR cognitive architecture, developed by Newell and colleagues, offered a parallel modeling tradition focused on unified theories of cognition that could both predict human problem-solving behavior and serve as a specification for intelligent system design. SOAR's chunking mechanism — by which practiced production-rule sequences become compiled into retrievable chunks — provided an explicit computational account of skill acquisition that informed decisions about how many practice trials to require and how to structure problem sequences to promote efficient chunking. [LE-LS-AP-015] John Anderson's ACT-R architecture, which drives the Cognitive Tutor family, was similarly derived from a detailed cognitive model of skilled performance, ensuring that the ITS's adaptive behavior was grounded in a validated account of how human learning actually operates. [LE-LS-PP-002]

Simulation for Learning System Design

Simulation in learning engineering ranges from cognitive simulation (using computational cognitive models to predict how different instructional sequences will affect learner performance before deployment) to virtual learning environment simulation (creating interactive digital environments that allow learners to practice skills that would be dangerous, expensive, or impossible to practice in real contexts). Chris Dede's research on immersive virtual environments established engineering parameters for using multi-user virtual environments and augmented reality in STEM education, demonstrating that carefully designed simulated environments can support complex inquiry skills and conceptual understanding not achievable through conventional instruction. [LE-LS-AP-006] Harvard's Graduate School of Education under Dede's leadership has been a primary site for developing and validating these simulation-based learning approaches. [LE-LS-PP-010]

The Human-Computer Interaction Institute at Carnegie Mellon extended simulation-based methods into the validation of learning system components, including assessment instruments, hint sequences, and interface layouts, enabling iterative refinement of these components before large-scale deployment. [LE-LS-CO-005] The Learning Engineering Toolkit treats simulation as an integral part of the evidence-based design cycle, noting that simulation-based testing — whether of cognitive models or interactive environments — allows teams to generate and evaluate design hypotheses at lower cost and higher speed than live deployment alone would permit. [LE-LS-GL-007]

¶ Subsections:

• 5.4.1 Role of Modeling in Learning Engineering
• 5.4.2 Simulation for Testing and Validation

From the Learning Engineering Toolkit

In the Learning Engineering Toolkit's chapter on engineering, the authors present modeling as one of the chief means by which engineers cope with everything they cannot fully know in advance, explaining that a model can act either as a pared-down stand-in for a complicated system or as a way to imitate systems too intricate or expensive to explore directly in the real world. [LET-04] They anchor the idea in everyday engineering: civil engineers draw on models of population growth and related demographics so that a solution is scaled appropriately for however long it will be in service, alongside models of material strength, weather, and the ways materials wear down and erode so that what they build endures. [LET-04] Some of these models are approximations rather than exact figures, carrying a cushion of allowable error that the engineer needs to grasp. [LET-04]

On this view, simulation is what enables engineers to try out and confirm designs that would be impractical to assess head-on. The Toolkit cites computational fluid dynamics models — highly detailed and accurate — used in aerospace to run tests and validation on new designs, offering the testing of wing shapes as an example. [LET-04] Models like these can preserve, and even introduce, complexities that would be hard or outright impossible to reproduce in reality. [LET-04] Turning back to learning, the authors note that learner models can serve to test learning systems or ideas, pointing to Piotr Mitros' team, which drew on models of learning such as the progression from passive to active and then constructive engagement. [LET-04]

The chapter also underscores that one system can be captured by multiple models at the same time. Models describe facets of a system and the mathematical relationships needed to understand and analyze it, and several models of the same system may be built to view it from different angles. [LET-04] For the edX platform, the authors illustrate this with separate models for learner interactions, the software architecture, the content within a course, and the data the system handles. [LET-04] The chapter's summary points capture the position plainly: systems get designed with the help of models spanning a range of fidelity. [LET-04]