Abstract
This paper introduces CARL, a computational model designed to learn abstract concepts through incremental analogical reasoning. Unlike traditional AI models that rely on static pattern matching between well-defined domains, CARL builds its understanding of a target domain (such as computer programming) by mapping causal structures from a familiar base domain (such as physical boxes) and iteratively debugging misconceptions. The research emphasizes that learning is an active process of extending analogies and correcting the errors that inevitably arise from them.
Introduction
In the early 1980s, AI research into analogical reasoning was dominated by models focused on structural similarity and partial pattern matching. However, Mark Burstein identified a critical limitation: human learners often approach a "target" domain with no prior representation. Furthermore, exhaustive bottom-up matching often leads to a combinatorial explosion of irrelevant data. Burstein’s model, CARL, addresses these issues by focusing on top-down causal abstractions and incremental learning through feedback.
The CARL Architecture
CARL (a system that learns via analogical reasoning) is built upon three fundamental pillars designed to mimic human cognitive processes:
- Top-Down Causal Mapping: Instead of matching every detail, CARL prioritizes the functional purpose of objects. For instance, when learning about variables through the analogy of a box, CARL focuses on the "containment" function rather than physical attributes like color or size.
- Abstraction Hierarchies and Virtual Relations: When a direct mapping is physically impossible—such as a number being "inside" a variable—CARL traverses a hierarchy of predicates to find a suitable abstraction. It generates a "virtual relation" (e.g.,
INSIDE-VARIABLE) that inherits properties from the base domain while adapting to the constraints of the programming environment. - Incremental Debugging: CARL treats learning as a hypothesis-testing cycle. It forms a model, tests it against new examples, and "debugs" its internal representation when errors occur.
Learning through Debugging
A key insight of the CARL model is that errors are not failures but essential data points. A common misconception for students learning BASIC is the "moving vs. copying" problem: assuming that an assignment statement like X = Y moves the value of Y into X, leaving Y empty—much like moving a ball between boxes.
CARL explicitly models this over-extension of an analogy. When the system observes that Y retains its value after the assignment, it triggers a debugging process. It searches for a related concept in its hierarchy that accounts for the persistence of the source value, eventually transitioning from a "physical movement" model to a "data replication" model. This incremental update allows the system to refine its understanding without discarding the useful parts of the original analogy.
Results and Implications
CARL successfully demonstrated the ability to navigate a sequence of BASIC programming lessons, evolving its understanding from concrete physical containment to abstract semantic representation. The system's ability to replicate and then correct common human student errors validated the "learning by debugging" approach.
While the model is teacher-dependent and relies on pre-defined predicate hierarchies, it remains a foundational contribution to cognitive architecture and Explainable AI (XAI). It highlights that robust learning requires the ability to reconcile the constraints of a new domain with the structures of the old, often through the iterative correction of analogical "leaks."
Summary updated