A brief comparison of cognitive tutors and cognitive assistants

In response to the reflection prompt for this week, I am focusing on the comparison between a cognitive tutor (Koedinger et al 2007) and a cognitive assistant (Le et al, 2018). Both are highly interactive learning tutors/agents but are starkly different in their designs and interaction capabilities, as well as their underlying design and theoretical rationales.

As described by Koedinger et al (2007), a cognitive tutor (CT) is domain-specific and offers a "rich, problem-solving environment with a variety of problem-solving tools and presents authentic problem-solving scenarios for the student to solve" (p 244). The modus operandi of this graphical interface is a thoughtful combination of feedback (and its timing), assessment, and forced navigation. The feedback is at a granular-level, that it, it is delivered at each step of the problem-solving process, rather than at the end of the process, and can be delivered as either feedback and/or, if requested by the student, hints. The CT takes a "mastery-knowledge" approach by tracking "knowledge components" such as key domain skills or concepts (p 246). For example, based on the number correct or incorrect answers, (or hints requested), the CT "maintains an estimate of the probability that the student has learned that knowledge component" and is ready to progress to the next level (p 247). The CT is "theoretically grounded" in the ACT-R theory of cognition, a model that can be described almost as a programming language based on model tracing (interpretation) and knowledge tracing (assessment) (p 248). One limitation in this, however, is that students might learn and ultimately be conditioned to think that asking questions is a 'bad thing' or holds back their progress. As well, the timing of the feedback, or the balance between withholding or offering feedback--the assistance dilemma--is still an open-ended research topic.

Fig. 1. Example of a cognitive tutor graphic interface (Koedinger et al)

Also highly interactive, cognitive assistants are "intelligent architectures and capabilities endowed with common sense knowledge and reasoning" that are meant to augment--not replace--human capabilities (Le et al, p 45). For example, Liza is a text-based pedagogical agent that is designed to foster reasoning skills and depends upon Natural Language Processing, Machine Learning, and for lack of a better term, data management. It consists of four steps (p 47):

1. Introduction (and partial customization)
2. Selection of 14 "stories" from seven training topics
3. Decision to offer hint, or not
4. Response evaluation and feedback

Each "story" is comprised of "text describing the premises and basic situation, a task question and a clarifying reformulated question, a category of heuristics/bias (e.g. sunk cost fallacy), a set of correct and incorrect values ..., a hint, an explanation and reactions to correct and incorrect answers" (p 47). One major limitation for an agent like Liza is the fact that humans can be (and often are) irrational and biased, so the training topics are limited to seven core elements of reasoning. Additionally, the agent uses probability to determine right or wrong answers (and the context in which these positive or negative tags occur) which could lead to erroneous determinations. Despite this limitation, the authors state that using a CA like Liza results in improved reasoning development over a non-interactive online course (p 50).

Although a CA and a CT are quite different, both are fascinating and rich in their implications for the learning sciences.

References

Goodfellow, I., McDaniel, P., & Papernot, N. (2018). Making machine learning robust against adversarial inputs. Communications of the ACM, 61(7), 56-66.

Koedinger, K. R., & Aleven, V. (2007). Exploring the assistance dilemma in experiments with cognitive tutors. Educational Psychology Review, 19(3), 239-264. 

Le, N. T., & Wartschinski, L. (2018). A Cognitive Assistant for improving human reasoning skills. International Journal of Human-Computer Studies, 117, 45-54.