TLDR: A new research paper by Paolo Liberatore introduces the concept of ‘abilities’ in iterated belief revision, shifting focus from what AI systems *must* do (postulates) to what they *can* achieve. It defines key abilities like plasticity, learnability, and amnesia, and analyzes existing revision mechanisms to show that different mechanisms possess different capabilities, suggesting a need for tailored approaches or combinations for diverse AI applications.
In the fascinating world of Artificial Intelligence, particularly within the realm of knowledge representation, how intelligent systems update their beliefs when faced with new information is a critical challenge. This process is known as belief revision. Traditionally, research in this area has heavily relied on a concept called ‘postulates’. These postulates are like a set of rules or constraints that define how a belief revision mechanism *must* behave. For example, a postulate might state that if new information is consistent with existing beliefs, it should be incorporated without causing other changes.
However, a recent research paper titled “Iterated belief revision: from postulates to abilities” by Paolo Liberatore introduces a fresh perspective. The paper argues that while postulates tell us what a revision mechanism *must* do, they often overlook what it *can* do. This is where the concept of ‘abilities’ comes in. Abilities describe the potential of a belief revision mechanism to reach certain states of belief or transform one state into another through a sequence of revisions.
Imagine a system learning about the world. Can it completely forget everything it once knew and start from scratch? Can it become absolutely certain about a single fact, dismissing all other possibilities? Can it make two previously distinct beliefs equally plausible? These are examples of ‘abilities’ that are crucial for real-world applications but are not fully captured by traditional postulates.
Also Read:
- Beyond Obedience: Why AI Needs Moral Responsibility, Not Just Compliance
- Navigating Accountability: Understanding Responsibility Gaps and Diffusion in AI Decision-Making
Understanding Key Abilities
The paper defines several important abilities:
- Plasticity: This is the ultimate flexibility, allowing a system to transform any belief state into any other non-flat (not completely ignorant) belief state.
- Learnability: The ability to learn from scratch, meaning a system can go from a state of complete ignorance (where all possibilities are equally likely) to any desired belief state. This is vital for systems starting with no prior knowledge.
- Amnesia: The capacity to completely forget all existing beliefs and return to a state of complete ignorance. This might seem counterintuitive but can be useful in scenarios where old information becomes entirely irrelevant.
- Equating: The power to make two previously distinct situations or beliefs equally plausible.
- Dogmatism: The ability to reach a state where one belief is held so strongly that all other possibilities are completely dismissed as impossible.
- Damascan: The ability to completely invert all beliefs, turning what was most believed into least believed and vice versa.
- Correcting: The ability to invert the order of belief between two specific situations.
- Believer: The ability to make a specific set of situations the most believed.
The research delves into various existing belief revision mechanisms, such as Natural, Lexicographic, Restrained, Very Radical, Full Meet, Severe, Moderate Severe, Deep Severe, and Plain Severe revisions. It then systematically analyzes which of these mechanisms possess which abilities. For instance, Natural, Lexicographic, and Restrained revisions are found to be learnable and damascan but not equating. On the other hand, Very Radical, Severe, Moderate Severe, and Deep Severe revisions are plastic and equating but not amnesic.
A significant finding is that no single existing revision mechanism possesses all desirable abilities. This highlights that different mechanisms are suited for different applications. For example, a mechanism that is good for learning from scratch might not be good for completely forgetting old information. The paper also points out that while some abilities might be missing in one mechanism, they can often be achieved by combining different mechanisms.
This work shifts the focus from merely defining how beliefs *must* change to understanding what complex transformations belief revision systems *can* achieve. It provides a valuable framework for selecting or designing belief revision mechanisms based on the specific needs and desired capabilities of an AI application. To explore the full details of this groundbreaking research, you can read the paper here.
