TL;DR
This paper introduces "Dialectical ASPIC+" (D-ASPIC+), a novel framework for structured argumentation that addresses the limitations of existing systems by ensuring rationality even when agents have limited computational resources. D-ASPIC+ achieves this by distinguishing between committed and supposed arguments, leading to a more realistic model of argumentation for real-world agents.
What Problem Does This Paper Address?
Structured argumentation frameworks, like the widely-used ASPIC+, are powerful tools for modeling reasoning and decision-making, especially in situations involving uncertainty or conflicting information. They represent arguments as structured entities built from knowledge bases and inference rules, and then evaluate their acceptability based on attacks and preferences. However, a critical limitation of current ASPIC+ systems is their implicit assumption of "unbounded resources." This means they assume agents can perform an infinite number of logical inferences, which is unrealistic for practical applications. This assumption leads to two main problems:
- Non-contamination: Inferences can be invalidated by adding syntactically disjoint but irrelevant information. This is like a minor detail derailing an entire logical deduction.
- Computational Feasibility: The requirement to generate all possible deductive inferences makes these systems computationally intractable for real-world agents.
This paper aims to bridge this gap by proposing D-ASPIC+, a framework that guarantees rationality (consistency, closure, and non-contamination) under realistic resource constraints.
Methods
The core innovation of D-ASPIC+ lies in its refined argument ontology. Unlike standard ASPIC+ arguments, D-ASPIC+ arguments are "upside down trees" where the "maximal fallible sub-arguments" (mfs) are differentially labeled. This means that for any given argument, its sub-arguments can be marked as either:
- Committed: These are arguments that the agent fully endorses and accepts.
- Supposed: These are arguments that the agent concedes for the sake of argument, perhaps to explore an interlocutor's position or to test a hypothesis.
This distinction allows D-ASPIC+ to satisfy rationality postulates without assuming logical omniscience. Specifically:
- Consistency and Closure: These are satisfied without assuming full deductive inferential capabilities or specific properties of preference relations.
- Non-contamination: The paper identifies conditions under which non-contamination is satisfied, focusing on the properties of preference relations and the available resources for constructing arguments. They propose that strict inference rules should be generated by proof theories that avoid "redundant" (syntactically disjoint) premises, which helps prevent contamination.
- Foreign Commitment Problem: D-ASPIC+ also addresses the "foreign commitment problem," where an agent might be forced to commit to an argument it doesn't actually endorse to challenge another argument.
The framework defines new notions of attacks and defeats that operate on these committed and supposed sub-arguments, ensuring that attacks primarily target committed information.
Datasets
Not discussed in the paper.
Key Insights and Ideas
- The Power of "Supposing": The most significant insight is that explicitly modeling "supposed" arguments, distinct from "committed" ones, is crucial for achieving resource-bounded rationality. This allows for more flexible and realistic dialectical exchanges.
- Deconstructing Arguments for Rationality: By focusing on the status of sub-arguments (committed vs. supposed), D-ASPIC+ can enforce rationality properties without requiring the entire argument to be fully reasoned out from scratch.
- Syntactic Disjointness as a Contamination Indicator: The paper highlights how syntactically disjoint information can be a key culprit in contamination, and proposes solutions by restricting the nature of strict inference rules.
- Bridging Theory and Practice: D-ASPIC+ represents a significant step towards making formal argumentation systems practical for real-world agents that operate under computational limitations.
Key Results
The paper establishes theoretical results demonstrating that D-ASPIC+ satisfies:
- Direct Consistency: For any admissible extension E, claims(E) will not contain contradictory or falsum statements, provided the arguments satisfy certain structural properties (P1, P2, P3).
- Closure under Strict Rules: If an argument can be strictly inferred from the claims of arguments in an extension, it will also be part of that extension.
- Non-Interference (a form of non-contamination): Under specific conditions on preferences and argument construction (P1-P4), expanding a theory with syntactically disjoint information does not alter the consequences of the original theory within the original symbol set.
Limitations and Future Work
The authors acknowledge that while D-ASPIC+ provides a framework for rationality under resource bounds, further work is needed.
- Proof Theory for Non-Contamination: A key direction is to formalize strict inference rules using proof theories that inherently avoid redundant premises. This would ensure non-contamination is satisfied regardless of preference relation properties.
- Computational Complexity: While the framework is designed for resource-bounded agents, the exact computational complexity of D-ASPIC+ reasoning remains an area for future investigation.
- Generalizing to Other Logics: The paper focuses on classical logic but suggests that the D-ASPIC+ approach can be extended to other non-monotonic logics.
The paper concludes by emphasizing that D-ASPIC+ offers a robust foundation for developing argumentation formalisms and non-monotonic logics that are both theoretically sound and practically applicable to resource-bounded agents.
Summary updated