Modeling in learning engineering serves as a formal mechanism for making design assumptions explicit, testable, and improvable. When a learning engineer builds a model — whether of learner knowledge states, domain structure, or system architecture — they are not merely documenting what exists but constructing a predictive framework that can be evaluated against observed data and refined when its predictions fail. This epistemological function of models — as hypothesis-generating and hypothesis-testing tools rather than passive descriptions — is central to what distinguishes learning engineering from less rigorous approaches to instructional design. [LE-LS-GL-007]
Conceptual and Domain Models
The most fundamental modeling task in learning engineering is domain modeling: constructing a formal representation of the knowledge or skill that learners are expected to acquire. At its most rigorous, this involves decomposing a target competency into discrete knowledge components — the units of knowledge or skill whose acquisition can be individually assessed and whose mastery is necessary for overall competence. Koedinger and Anderson's 1990 analysis of expert geometry problem-solving provided the methodological template for this process, showing how expert behavior could be analyzed into discrete production rules that specified exactly what knowledge was being applied at each step of a problem solution. [LE-LS-AP-001] This granular domain model served as the specification for the Cognitive Tutor's assessment and hint systems, ensuring that the system's instructional behavior was precisely aligned with the structure of the target knowledge.
The SOAR architecture provided a complementary modeling framework focused on unified theories of cognitive processing rather than domain-specific skill structure. By representing all learning as a form of chunking — the compilation of practiced processing sequences into retrievable higher-level units — SOAR offered a single computational account of how both novice and expert performance emerge from the same underlying mechanism. [LE-LS-AP-015] This modeling approach has direct implications for curriculum design: it suggests that problem sequences should be structured to promote chunking through controlled variation and repetition, and that mastery criteria should be calibrated to the point at which target chunks are reliably formed rather than simply to a fixed number of correct responses. Kenneth Koedinger's empirical research on learning curves — analysis of how error rates decline across problem-solving practice — provided validation methods for assessing whether domain models accurately captured the structure of target skills. [LE-LS-PP-005]
Student Models and Learner State Representation
Student models represent the current knowledge state of individual learners, enabling adaptive systems to customize instruction in response to what each learner knows and can do rather than treating all learners identically. Bayesian Knowledge Tracing, the foundational student modeling approach, uses a hidden Markov model structure in which the true knowledge state (learned or not-yet-learned) is treated as a latent variable inferred from observable responses. [LE-LS-AP-011] The model's four parameters — prior probability of mastery before instruction, probability of a correct response from a learner who has mastered the skill (correctness given mastery), probability of an incorrect response despite mastery (slip rate), and probability of a correct response without mastery (guess rate) — can be estimated from historical data and then applied to new learners in real time.
The full Cognitive Tutor architecture integrated the student model with the domain model and a pedagogical model that specified how to select problems and generate hints in response to the student model's estimates. [LE-LS-AP-002] John Anderson's ACT-R cognitive architecture provided the theoretical foundation linking these modeling components: ACT-R's production rule firing mechanism is the computational substrate that the Cognitive Tutor's domain model encodes, ensuring that the system's behavior is grounded in a validated account of human cognitive processing rather than engineered heuristics disconnected from cognitive science. [LE-LS-PP-002] This integration of domain model, student model, and cognitive architecture — each validated empirically and connected to the others through explicit theoretical links — represents the gold standard for modeling in learning engineering practice. [LE-LS-GL-007]