Introduction
Published in 2001, this paper provides a comprehensive survey and evaluation of how Declarative Logic Programming (DLP)—specifically Answer Set Semantics—can serve as the "thinking component" for intelligent software agents navigating the World Wide Web. The authors argue that the messy, heterogeneous nature of web data requires agents with formal reasoning capabilities to act as effective mediators between users and information sources.
The Problem: Data Mediation in a Heterogeneous Web
In the early 2000s, the Web was a "Wild West" of unstandardized data. For an automated agent to be useful, it couldn't just fetch pages; it had to act as a mediator. This involves several complex logical tasks:
- Decomposing a user's complex request into sub-queries for different sources.
- Selecting the most reliable sources based on meta-knowledge.
- Merging inconsistent data (e.g., two different sites reporting different prices for the same item).
- Updating internal knowledge as the world changes.
Most systems at the time relied on ad-hoc procedural code (if-then-else blocks), which the authors argue is brittle and hard to maintain. They propose using DLP to create a modular "Thinking Component" that allows agents to reason about data using formal logic rather than hard-coded procedures.
The Modular Agent Architecture
A key contribution of the paper is the advocacy for a strict architectural separation between an agent's communication layers and its reasoning engine. This modularity allows the reasoning engine to be updated or swapped without rebuilding the entire agent infrastructure.
As illustrated in the architecture above, the Information Agent sits between the User Interface and the Data Sources. Its core is the Thinking Component, which houses Domain Knowledge and Meta-Knowledge about Sources. This separation ensures that the logic governing how data is interpreted is distinct from the mechanics of how it is retrieved and presented to the user.
Dimensions of Logical Reasoning
The paper categorizes and evaluates DLP techniques across four critical dimensions essential for agent intelligence:
- Priority Handling: Methods for deciding which rules or data sources "win" when they contradict each other (e.g., preferring a primary source over a secondary one).
- Knowledge Revision and Update: Techniques for an agent to incorporate new information. The authors distinguish between revision (refining knowledge about a static world) and update (reacting to a world that has physically changed).
- Quantitative Reasoning: Incorporating "fuzzy" concepts like costs, weights, and probabilities into rigid logical frameworks to handle uncertainty.
- Temporal Reasoning: Allowing agents to reason about the sequence of actions and how their knowledge base evolves over time.
Key Insights and Results
The authors identify Answer Set Semantics as the most promising foundation because it handles "negation as failure" (reasoning about what we don't know) better than classical logic.
A significant insight is the "Conflict Paradox": when merging data from the Web, inconsistency is a feature, not a bug. DLP methods like "Weak Constraints" allow an agent to find the "best possible" answer even when the data is contradictory. Furthermore, the paper highlights that priorities should be dynamic; an agent should be able to reason that "Source A is more reliable for sports, but Source B is better for finance."
The paper provides a qualitative evaluation of existing frameworks like IMPACT, HERMES, and COIN. It concludes that while many handle "static" priorities well, very few can handle "dynamic" updates where the rules of the game change mid-stream.
Limitations and Future Work
Despite the power of DLP, the paper identifies several open research frontiers:
- Unified Frameworks: The lack of a "Grand Unified Theory" that combines priorities, updates, and temporal reasoning into one package.
- XML Integration: Using emerging standards like XML as a candidate for declarative merging strategies.
- Computational Complexity: Many powerful logic methods are NP-hard, making real-time performance a significant hurdle for web-scale agents.
- Meta-Reasoning: The need for better ways for agents to reason about their own reasoning process without falling into logical infinite loops.
In summary, while the building blocks for rational agent behavior exist within logic programming, integrating them into a single, efficient framework remains a significant challenge for the field of intelligent information systems.
Summary updated