TL;DR
This paper argues that domain-dependent control information in AI reasoning systems is unnecessary and can be replaced by domain-independent control rules applied to domain-dependent modal facts. This reframing simplifies AI architectures, eliminates the need for "metametalevels" of reasoning, and shifts the focus of learning towards acquiring richer domain knowledge.
What Problem Does This Paper Address?
In Artificial Intelligence, particularly in declarative systems, reasoning processes are guided by control information. This information dictates how the system should search for solutions. Traditionally, this control information has often been domain-specific, meaning it's tailored to the particular knowledge base or problem domain (e.g., "when reasoning about countries, prioritize checking for invasions before alliances"). The authors observe that this domain-dependency is problematic and leads to complex architectures, including the need for "metametalevels" to manage control strategies themselves. They question whether this specificity is truly necessary and propose a cleaner, more generalizable approach.
Methods
The core of the paper's argument rests on a distinction between two types of "metalevel" knowledge:
- Modal Knowledge: This is knowledge about the declarative knowledge itself, describing its structure or state, but not dictating actions. Examples include "I know that Iraq invaded Kuwait" or "The database contains a fact about alliances."
- Control Knowledge: This is information that tells the system what to do with its knowledge, guiding the search process.
The authors claim that any domain-dependent control rule can be decomposed into:
- A domain-independent control rule: This rule is general and applies across different domains.
- A domain-dependent modal fact: This fact describes the specific structure of the current domain's knowledge base, indicating when the domain-independent rule can be applied.
They demonstrate this by showing how a domain-specific control rule for a "hostile country" example can be replaced by a general rule that checks for a "type" predicate (a modal fact) describing the structure of the database. This "type" predicate is derived from the database's content and is essentially a generalization of the domain-specific structure.
Datasets
Not discussed in the paper. The paper uses illustrative examples of declarative databases (e.g., country hostility, geometric properties) to demonstrate its theoretical points, rather than relying on formal datasets.
Key Insights and Ideas
- Decomposition of Control: The most significant insight is that domain-dependent control rules are often a conflation of a general control principle and a specific observation about the domain's knowledge structure. By separating these, we achieve greater generality.
- Elimination of Metametalevels: The proposed decomposition removes the need for higher-order reasoning about control strategies. Since control rules are domain-independent, they don't need to be adapted or managed based on the domain.
- Shift in Learning Focus: If domain-dependent control rules are not learned, learning efforts should instead focus on acquiring richer base-level or modal information. This information can then be leveraged by the existing domain-independent control rules.
- Reinterpreting Existing Knowledge: Many domain-specific rules can be reinterpreted as instances of general principles applied to specific modal facts. For example, the "frame axiom" (that achieved subgoals remain satisfied) can justify domain-independent control rules for planning.
- Benign World Assumption: The authors suggest that the reason domain-independent control principles are often sufficient is that our world is "benign enough" to allow for broadly applicable rules.
Key Results
The paper presents a theoretical argument and a formal proposition (Proposition 2.1) stating that any domain-dependent control rule can be replaced by a domain-independent control rule and a modal sentence describing the search space structure. The computational cost of this replacement is minimal, equivalent to a single inference step, especially if modal information is cached.
Qualitatively, the paper argues that this approach leads to:
- Simpler AI architectures.
- More robust and generalizable control strategies.
- A clearer understanding of how learning should proceed.
- A unified view of various AI reasoning techniques.
Limitations and Future Work
- Practical Implementation: While the theoretical argument is strong, the paper acknowledges that the "general construction is extremely weak" and points to Section 3 for a more powerful application. The practical implementation and efficiency of caching modal information are mentioned as areas for further consideration.
- Learning New Control Rules: The paper suggests that learning should focus on base-level or modal information, implying that the development of new control rules might be less critical. However, it also notes that extracting general principles from existing domain-specific rules is an ongoing research area.
- Experimental Validation: The authors admit that the claim that domain-independent control rules are generally useful is an "experimental question." They suggest that further research is needed to validate this observation across a wide range of problems.
- Extending Declarative Languages: The paper hints at the need to extend declarative languages to handle probabilistic or default information, which could further enhance the power of domain-independent control principles.
Summary updated