Magazine article AI Magazine

Extending the Diagnostic Capabilities of Artificial Intelligence-Based Instructional Systems

Magazine article AI Magazine

Extending the Diagnostic Capabilities of Artificial Intelligence-Based Instructional Systems

Article excerpt

Computer-based educational technology has had a transformative impact in every imaginable educational context - from tablet-based educational games for preschoolers, to massively open online courses for university students and independent learners. Artificial intelligence (AI) research solutions have the potential to boost the impact of these systems by enhancing their diagnostic capabilities. For example, unlike one-on-one human tutoring, it is rare to find computer-based learning environments that not only provide opportunities for practice on complex problems (such as working on multistep algebra problems, or practicing programming by writing multiline code), but also provide contextually relevant feedback and guidance based on an analysis of individual problem-solving actions. In contrast, because of the technical challenge of making sound diagnostic inferences, we typically find two extremes represented in the design of instructional systems: either systems that constrain the learner by restricting allowable actions during skill acquisition (for example, multiple choice problems) or systems that provide opportunities for open-ended practice, with the only corrective feedback coming from the natural feedback inherent to the task environment (for example, a programming problem that a student solves on his or her own - with built-in feedback from the interpreter or compiler). The problem with the first extreme is that students develop skills in context and complexity that are different from their eventual application context. The problem with the latter is that students may flounder with inadequate feedback, or have incorrect or superficial strategies reinforced in the absence of corrective feedback and guidance.

These challenges point to the need for diagnostic techniques that are (1) tied to problem solving - capable of interpreting and assessing individual actions in the context of relatively unconstrained problem-solving activities (for example, writing a computer program, flying a scenario in a flight training simulator); (2) automated - reducing the need for manual intervention; (3) objective - minimizing confounds stemming from subjective biases; and (4) fine-grained - providing feedback at the level of individual problem-solving actions, as opposed to problem-solving outcome alone.

Cognitive Tutors: A Promising AI Approach

Cognitive tutors, based on the adaptive control of thought - rational (ACT-R) cognitive architecture, represent one promising approach for boosting the diagnostic capability of interactive learning environments (Anderson et al. 1995). The development of the ACT-R theory of cognition, which describes how humans perceive, think, and act, has been led by psychologist John Anderson and colleagues at Carnegie Mellon University (Anderson and Lebiere 1998). ACT-R is instantiated in the form of a programming language with primitive constructs that embody specific assumptions about human cognition. It has been used to model human performance in a broad array of complex cognitive tasks, ranging from automobile driving (Salvucci 2001) and tactical decision making (Anderson et al. 2011) to algebra and computer programming (Corbett, Anderson, and O'Brien 1995).

Cognitive model-based diagnosis is particularly well suited for environments where learners acquire skills through hands-on practice. In many complex problem-solving domains, learners have access to a broad range of problem-solving actions (operators) that can be combined to transform a problem from some initial state to a goal state through a set of intermediate problem states. In complex task domains, the number of actions students have at their disposal can be quite large. By interacting with elements of the learning environment, these actions can produce a vast set of possible intermediate problem-solving states. The primary task of the novice, learning to navigate unfamiliar problem spaces, is to reach goal states efficiently through one or more possible sequences of states. …

Search by... Author
Show... All Results Primary Sources Peer-reviewed

Oops!

An unknown error has occurred. Please click the button below to reload the page. If the problem persists, please try again in a little while.