To understand why a machine learning system might fail, we first need to understand how it generalizes. In traditional similarity-based learning, a model looks at hundreds of examples of a concept—say, a coffee mug—and statistically infers that a flat base and a handle are important features.
By contrast, Explanation-Based Learning (EBL) and related analytical frameworks operate on a different principle: they learn from a single example. If you show an EBL system a single coffee mug, it doesn't just record the pixels or features. Instead, it uses a pre-existing body of knowledge—a domain theory—to causally explain why the object functions as a mug. By applying known rules of physics (e.g., liquid requires containment, gravity pulls objects downward), the system deduces that the flat bottom prevents tipping and the handle prevents burning. Once a valid explanation is constructed, the system strips away the irrelevant details (like the color or material) and creates a generalized schema for future use.
This paradigm, pioneered by researchers like Gerald DeJong in foundational papers such as Generalizations Based on Explanations and Acquiring Schemata Through Understanding and Generalizing Plans, offers incredible sample efficiency. However, it introduces a massive hidden cost: the domain theory dependency.
The system can only learn what it can already explain. This dependency shifts the AI bottleneck from data collection to knowledge engineering, creating a strict boundary on what an analytical system can actually acquire.
The Incompleteness Problem: Total Learning Paralysis
The most immediate consequence of the domain theory dependency is that an analytical learner cannot truly learn from scratch. As Gerald DeJong observed in Automatic Schema Acquisition in a Natural Language Environment, the breadth of learnable concepts is strictly limited by the depth of the initial knowledge base.
If the domain theory is incomplete, the system suffers from total learning paralysis. Consider an agent trying to learn a new narrative structure or a physical procedure. If the system lacks the foundational axioms required to bridge a cause and an effect, the explanation mechanism breaks down entirely:
- In Natural Language: In Learning Schemata for Natural Language Processing, Raymond J. Mooney demonstrated that without a comprehensive internal model of basic human goals and physical causality, an NLP system cannot generalize new conceptual schemata from a text. It cannot understand why a character acts, so it cannot learn the underlying pattern of the behavior.
- In Physical Tasks and Robotics: Alberto Maria Segre argued in On the Operationality/Generality Trade-off in Explanation-based Learning that if a system's internal model of physics or kinematics cannot account for the real-world complexity of a task, it becomes impossible to generate valid explanations or learn new schemas from physical demonstrations.
- In Expert Domains: This problem is acute in specialized fields. Tom M. Mitchell et al. showed in LEAP: A Learning Apprentice for VLSI Design that analytic learning systems are completely blocked in domains where the underlying causal or logical principles are not yet fully codified.
If you cannot formally prove why an expert demonstration works, you cannot generalize it (Sridhar Mahadevan in Verification-based Learning: A Generalized Strategy for Inferring Problem-Reduction Methods). Similarly, if an agent cannot explain why a plan failed because its environmental axioms are incomplete, it cannot derive the constraints necessary to prevent that failure in the future (Ajay Gupta 0005 in Explanation-based Failure Recovery).
The Flawed Theory Problem: Misguided Learning
What happens if the domain theory isn't just incomplete, but slightly incorrect, inconsistent, or weak? This is where the dependency becomes dangerous. Instead of simply failing to learn, the system actively misguides itself, deriving highly confident but fundamentally incorrect rules.
Francesco Bergadano framed this beautifully in The Problem of Induction and Machine Learning as a search space problem. Induction requires searching through an exponential space of possible hypotheses. A correct domain theory acts as a vital filter, narrowing down the search space to predictive rules. However, if the underlying theory is flawed, the system may permanently exclude the true rule from its search space, rendering the system incapable of correct induction.
Correct Theory: [Possible Hypotheses] ---> Filters down to ---> [True Rule]
Flawed Theory: [Possible Hypotheses] ---> Filters down to ---> [Incorrect Rule] (True Rule Excluded)
This vulnerability manifests across several learning architectures:
- Feature Relevance Errors: Andrea Danyluk showed in The Use of Explanations for Similarity-based Learning that when symbolic rule bases are inconsistent or sparse, the generated explanations assign incorrect relevance to features, completely derailing downstream similarity matching.
- Deductive Initialization Traps: When a system uses a weak domain theory to initialize its search paths, it can be forced down incorrect logical trajectories that are mathematically difficult to recover from during later concept induction (Marco Botta et al. in SMART+: A Multi-Strategy Learning Tool).
- Suboptimal Default Rules: In non-monotonic reasoning frameworks, incomplete or biased background knowledge causes the system to generate flawed default rules and incorrect hierarchical exceptions (Katsumi Inoue et al. in Learning Extended Logic Programs).
This is not merely a historical issue for symbolic AI; it persists in modern neuro-symbolic and weakly supervised setups. For example, Zelin Xu et al. demonstrated in Spatial-Logic-Aware Weakly Supervised Learning for Flood Mapping on Earth Imagery that injecting spatial logic as a prior constraint into deep networks introduces a strict dependency on knowledge base quality. If those spatial logic rules are flawed or incomplete, they actively bias neural models toward physically inaccurate results.
The Structural Boundaries of Generalization
Even when a domain theory is perfectly accurate, the dependency imposes rigid structural limits on how far a concept can be generalized. Because the system's generalization is derived directly from its explanation trace, the system is trapped within the boundaries of its initial proof structure.
As Jude W. Shavlik et al. detailed in BAGGER: An EBL System that Extends and Generalizes Explanations, systems that rely strictly on existing explanations are passive; they cannot perform independent searches to fill logic gaps or synthesize entirely new rules if the initial trace lacks information.
This structural rigidity creates significant barriers in three key areas:
1. The "Generalizing to N" Problem
Standard EBL is excellent at replacing specific constants with variables (e.g., learning that a rule for "Block A" applies to "any Block $x$"). However, it struggles to generalize structure—such as learning a concept that involves an arbitrary number of steps or objects ($N$). Jude W. Shavlik et al. showed in An Explanation-based Approach to Generalizing Number that scaling an explanation to an arbitrary number of entities requires highly specialized, recursive background knowledge. Without it, the system cannot bridge the gap between a single instance and a scalable rule.
2. The Bootstrapping Barrier
Because analytical systems require a pre-existing causal framework to evaluate data, they cannot easily bootstrap theories in entirely novel environments. Michael J. Pazzani et al. argued in The Role of Prior Causal Theories in Generalization that this dependency on predefined generalization heuristics and foundational knowledge creates a severe barrier to entry for truly autonomous learning. If an agent doesn't have an internal model of how primitive actions affect a new environment, it is forced to abandon analytical generalization entirely and fall back on rote memorization (Scott B. Huffman et al. in Learning Procedures from Interactive Natural Language Instructions).
3. The Abstraction Mismatch
The utility of any learned rule is tightly bound to the abstraction level of its background knowledge. If the primitive concepts defined for an inductive learner are too granular or poorly aligned with the domain, the resulting rules fail to provide meaningful insights (Nico Jacobs et al. in Inductive verification and validation of multi agent systems). Furthermore, manually constructing these formal, highly structured domain theories for complex, natural datasets (like computer vision) remains a monumental challenge (Qiang Sun et al. in Explanation-Based Learning for Image Understanding).
Summary
Explanation-Based Learning achieves its remarkable sample efficiency by trading data for knowledge. It works because it uses a domain theory to uncover the causal "why" behind a single example. Yet, this very mechanism reveals why the idea is so deeply constrained: the system's capacity to learn is fundamentally locked within the completeness, accuracy, and abstraction level of the knowledge it already possesses.
Key Foundational Work
Generalizations Based on Explanations (1981) — contributed 1 concept elements to this direction.
Automatic Schema Acquisition in a Natural Language Environment (1982) — contributed 1 concept elements to this direction.