TL;DR
This paper introduces a probabilistic extension to the stit logic framework, enabling the formal modeling of "attempts" where actions might fail. By incorporating subjective probabilities into the logic, it allows agents to reason about their beliefs regarding other agents' choices and their own chances of success, moving beyond the traditional assumption of guaranteed action success.
What Problem Does This Paper Address?
Traditional formalisms for modeling agency, like stit logic, often assume that when an agent "sees to it" that something happens, it is guaranteed to happen. This is unrealistic, as actions can fail due to unforeseen circumstances or the actions of other agents. The paper addresses this gap by developing a framework that can model situations where an agent's choice might not lead to the desired outcome. This is crucial for understanding concepts like responsibility and liability, where the success or failure of an action is central. The authors also highlight the need to formally distinguish between an "attempt" and a guaranteed action, as an attempt implies a possibility of failure and a comparative notion of maximizing chances of success.
Methods
The core contribution is the development of XSTITP, an extension of the XSTIT logic. This extension incorporates subjective probabilities into the object language, allowing agents to express beliefs about the choices of other agents. The semantics are based on "h-effectivity functions" which map states and histories to sets of possible next states, specialized to be history-relative.
Key methodological innovations include:
- Probabilistic Semantics: Introducing subjective probabilities (B) that represent an agent's beliefs about other agents' choices.
- Chance of Success (CoS) Function: Defining a function that quantifies an agent's belief about the probability of achieving a certain outcome given their choice and beliefs about others.
- Attempt Operator ([{ag} xatt]): Defining a new modal operator that captures the notion of an "attempt" as a choice that is optimal in the sense of maximizing the chance of success, relative to other available choices. This operator includes a crucial "side condition" that requires there to be at least one alternative choice with a strictly lower chance of success.
The paper also provides a Hilbert-style axiomatization for the base probabilistic logic (XSTITP) and discusses its completeness.
Datasets
Not discussed in the paper.
Key Insights and Ideas
- Subjective Probabilities for Beliefs: The paper demonstrates how subjective probabilities can effectively model an agent's beliefs about the simultaneous choices of other agents, which is fundamental to understanding action failure.
- Attempt as a Comparative Notion: The formalization of "attempt" as a choice that maximizes the chance of success, rather than just having a non-zero chance, is a key insight. This highlights that an attempt is judged relative to other available options.
- Distinction Between Choice and Action: The authors emphasize that an agent can be certain about its own choice but mistaken about the choices of others, leading to potentially unsuccessful actions. This distinction is crucial for modeling agency realistically.
- Independence of Agency: The probabilistic framework allows for a richer notion of independence, where agents' choices can be multiplied to determine the probability of joint outcomes, offering a new dimension beyond traditional stit logic.
Key Results
- XSTITP Logic: A new probabilistic extension of stit logic is formally defined with its semantics and a Hilbert-style axiomatization.
- Chance of Success Operator: A probabilistic stit operator,
[{ag} xstit=c]φ, is introduced, allowing statements like "agent ag sees to it that φ with at least probability c." - Attempt Operator: A formal definition for the "attempt" operator,
[{ag} xatt]φ, is provided, capturing the idea of an optimal choice in terms of maximizing success probability. - Validities: Several valid formulas are presented, demonstrating that the probabilistic stit operator generalizes the standard stit operator and that the attempt operator exhibits desirable properties like independence and the inability to attempt both a proposition and its negation simultaneously.
Limitations and Future Work
- Axiomatization of Probabilistic Operators: The paper acknowledges that axiomatizing the probabilistic stit and attempt operators is expected to be difficult, especially for the probabilistic stit operator due to probabilities being in the object language.
- Group Choices: A significant area for future work is the generalization of the theory to model group choices. This would involve exploring how information is pooled within a group and how agents' beliefs about other group members' choices might differ from their beliefs about external agents.
- Modeling "Deliberate" Choices: The "side condition" in the attempt operator hints at a connection to "deliberative" choices, which could be further explored.
Summary updated