TL;DR
This paper argues that ranked preferential models (RPMs), a popular framework for non-monotonic reasoning and belief revision, are insufficient for handling iterated belief change. Specifically, they fail to guarantee the "reversibility" of belief change, meaning we can't always undo a belief update by applying a counteracting update. This highlights the need for richer models that incorporate numerical strengths of belief.
What Problem Does This Paper Address?
Non-monotonic reasoning is crucial for artificial intelligence because it allows systems to make plausible inferences that can be retracted in light of new information. Traditional approaches to this, like default logic and circumscription, have faced theoretical challenges. Ranked preferential models (RPMs) emerged as a promising solution, offering a more principled way to model belief revision using a ranking of possible worlds. However, the paper questions whether RPMs are truly adequate, particularly when dealing with sequences of belief changes. The core problem is understanding how beliefs should evolve over time, especially when we need to "unlearn" something we previously accepted.
Methods
The paper first establishes the connection between non-monotonic inference and belief revision, referencing Gärdenfors' axioms for belief revision. It then delves into the semantics of RPMs, explaining how they represent belief states as well-ordered partitions of possible worlds. The central argument is built by introducing two key axioms for iterated belief change:
- (B9) Precision: If we learn a more specific piece of information (B entails A), the resulting beliefs should be the same as if we had learned only the more specific information initially.
- (B10) Conflict Resolution: If we learn two inconsistent pieces of information (A and B), the later information effectively "wipes out" the earlier information.
The paper then demonstrates that RPMs, while satisfying these axioms for single belief changes, fail to guarantee reversibility when these axioms are applied iteratively. The authors also explore whether numerical strengths of belief could salvage RPMs, but conclude that this approach also falls short.
Datasets
Not discussed in the paper.
Key Insights and Ideas
- RPMs are not enough for iterated belief change: While RPMs provide a good model for single belief updates, they struggle when multiple updates occur sequentially, especially when those updates need to be undone.
- Reversibility is a crucial property: The ability to "unlearn" information is a fundamental aspect of rational belief revision. If a system cannot reverse a belief change, it's a significant limitation.
- The "wiping out" effect of new information: Axiom (B10) formalizes the intuition that learning something that contradicts prior beliefs effectively discards the prior belief. This is a key mechanism that RPMs struggle to manage in an iterative context.
- Numerical strengths of belief might be necessary: The paper strongly suggests that qualitative models like RPMs, without numerical degrees of belief, are inherently limited in handling complex belief dynamics.
Key Results
The paper provides a formal proof demonstrating that RPMs, even when augmented with axioms (B9) and (B10), cannot guarantee the reversibility of belief change. This means that for certain sequences of belief updates, there's no way to return to a previous belief state by applying a counteracting update, even if that counteracting update is based on the original information. The authors illustrate this with a specific example involving four worlds and two different rankings, showing how a belief update can lead to a state from which it's impossible to recover the original rankings.
Limitations and Future Work
The authors acknowledge that their proposed axioms (B9 and B10) are not claimed to be complete for iterated belief change. A significant limitation of RPMs, as highlighted by the paper, is their inability to handle reversibility without numerical strengths of belief. The paper suggests that future research should explore representation theorems that characterize belief revision functions satisfying Spohn's axioms (which do incorporate numerical uncertainty) and investigate how numerical uncertainty can be integrated into reasoning frameworks.
Summary updated