💻 computer science

From Actions to Obligations: A Deontic Action Model Logic

This paper introduces Deontic Action Model Logic (DAML), a novel dynamic modal framework that extends Action Model Logic to formally derive context-sensitive obligations in multi-agent systems by evaluating actions based on their expected deontic value, and establishes its soundness, completeness, and expressive power through axiomatization and case studies like the Miners' Puzzle.

Original authors: Giorgio Cignarale

Published 2026-05-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Giorgio Cignarale

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are standing at a crossroads in a foggy forest. You have a map, but it's incomplete. You know there are two paths ahead, but you don't know which one leads to a safe village and which one leads to a cliff. You also know that your friends are with you, and they might know more than you do.

This paper introduces a new "rulebook" (a logic system) called DAML (Deontic Action Model Logic) to help agents (like robots, software, or even people) make the right choice when they are unsure of the facts.

Here is how the paper breaks it down, using simple analogies:

1. The Problem: "What should I do?" vs. "What do I know?"

Usually, logic helps us figure out what is true (Epistemic Logic) or what is good (Deontic Logic).

Epistemic Logic is like a detective asking, "What do I know for sure?"
Deontic Logic is like a judge asking, "What is the right thing to do?"

The problem is that in real life, you often have to make a "right" decision without knowing all the facts. The paper argues that to figure out what you ought to do, you need to combine these two questions. You can't just look at the rules; you have to look at the rules through the fog of your uncertainty.

2. The Solution: The "Expectation Score"

The authors propose a way to calculate a "score" for every possible action you could take. Think of it like a video game where every outcome has a point value:

Desirability: How good is the outcome? (e.g., Saving 10 lives = 100 points; Saving 0 lives = 0 points).
Likelihood: How likely is this outcome to happen based on what you currently know?

The logic calculates an Expected Deontic Value. It's like averaging the points of all possible futures you can imagine. If Action A leads to a 50% chance of 100 points and a 50% chance of 0 points, your score is 50. If Action B leads to a 100% chance of 40 points, your score is 40.

The Rule: You are "obligated" to choose the action with the highest expected score.

3. The Engine: "Action Models" (The Simulation Machine)

How does the system handle the "fog"? It uses something called Action Models.
Imagine you are a chess player. Before you move a piece, you mentally simulate: "If I move here, what happens? If my opponent moves there, what happens next?"

In this paper, the logic does exactly that. It takes your current "map" (what you know) and runs a simulation for every action you could take.

It creates a "mini-world" for every choice.
It checks the score of each mini-world.
It compares them to see which one is the "best" bet.

4. Real-World Examples from the Paper

Example A: The Miners' Puzzle

The Scene: Ten miners are trapped in either Shaft A or Shaft B. You don't know which one. You have sandbags to block one shaft, but not both.
The Dilemma:
- If you block Shaft A and they are there, everyone lives (10 points). If they are in B, they all drown (0 points).
- If you block Shaft B and they are there, everyone lives (10 points). If they are in A, they all drown (0 points).
- If you block neither, the water rises halfway, and only the lowest miner dies (9 points saved).
The Logic's Verdict: Even though blocking a shaft might save everyone, you don't know which one to block. The "average" score of blocking a shaft is low because of the risk of killing everyone. The "average" score of blocking nothing is high because it guarantees saving 9 people.
The Result: The logic says you ought to block neither shaft. It's the safest bet given your lack of information.

Example B: The Doctor and the Nurse

The Scene: A patient needs medicine. Drug D is usually perfect, but if the patient is allergic, it kills them. Drug D' is okay for everyone.
The Characters:
- Alice (Doctor): Wants to give the best drug but doesn't know the patient is allergic.
- Bethany (Nurse): Knows the patient is allergic and knows Alice doesn't know.
The Logic's Verdict:
- From Alice's perspective (who doesn't know), she thinks giving Drug D is the best move.
- From Bethany's perspective, she knows that if Alice knew the truth, she would choose Drug D'.
- Therefore, Bethany is obligated to tell Alice the truth. Why? Because by telling Alice, she changes Alice's "knowledge map," which changes Alice's "expected score," leading Alice to choose the safe drug.

5. Why This Matters

The paper proves that this system is mathematically sound (it doesn't break) and complete (it can solve any problem within its rules).

It provides a formal way for computers or agents to say: "I don't know everything, but based on what I do know and the rules of right and wrong, this is the most responsible action I can take."

In short: The paper builds a mathematical calculator that helps you make the most ethical decision possible, even when you are flying blind. It turns "I think I should do this" into a rigorous, provable fact based on your current knowledge.

Technical Summary: From Actions to Obligations: A Deontic Action Model Logic

Problem Statement

Deontic logic, which studies normative concepts like obligations and permissions, has traditionally struggled to integrate epistemic considerations (what agents know or believe) with dynamic action modeling. While extensions like STIT logic address agency and choice, and dynamic deontic logics evaluate obligations relative to actions, these frameworks often lack explicit mechanisms for reasoning under uncertainty. Conversely, Dynamic Epistemic Logic (DEL) and Action Model Logic (AML) excel at modeling information updates but lack formal tools for normative evaluation.

A critical gap exists in modeling scenarios where agents must select actions under incomplete information to satisfy normative constraints. A prime example is the "Miners' Puzzle," where the optimal action depends on the desirability of outcomes weighted by their likelihood, a concept not naturally captured by standard deontic modalities. Existing frameworks often fail to provide a principled account of conditional and context-sensitive obligations in multi-agent systems where agents must compare the expected deontic value of alternative actions.

Methodology

The paper introduces Deontic Action Model Logic (DAML), a dynamic modal framework that synthesizes Action Model Logic (AML) with a "Bayesian" semantics for deontic modals, inspired by the work of Lassiter. The methodology proceeds through the following steps:

Semantic Foundation:
- The framework utilizes Graded Kripke Models ( $M = \langle W, R_i, V, f \rangle$ ), where a desirability function $f: W \to \mathbb{N}$ assigns a moral or practical value to each possible world.
- Action Models ( $U = \langle E, Q_i, pre \rangle$ ) are employed to simulate the epistemic and physical consequences of agents' actions via the product update operation ( $M \otimes U$ ).
- To handle uncertainty, the logic defines Agent-based Action-Generated Submodels. These are submodels of the updated Kripke model constructed from the perspective of a specific agent $i$ and a specific action $\alpha_i$ .
Deontic Expectation:
- The core mechanism is the Deontic Expectation Function ( $E_i$ ), which calculates the expected deontic value of an action. This is computed as the average of the desirability values ( $f(w)$ ) of all worlds accessible to agent $i$ within the action-generated submodel, assuming a uniform probability distribution over accessible worlds.
- An action is deemed "best" if it yields the highest expected deontic value among the available alternatives at a decision point.
Logical Language and Semantics:
- The language $L_{DAML}$ $L_{D A M L}$ extends AML with two new elements:
  - Expectation Atoms ( $e^{\alpha_i}_i$ ): Atomic propositions that are true if the expected deontic value of action $\alpha_i$ is greater than or equal to that of any other available action.
  - Obligation Modality ( $O_i(U, \alpha_i | \phi)$ ): A binary modality reading "Agent $i$ ought to perform action $\alpha_i$ under condition $\phi$ ."
- Truth Conditions: $O_i(U, \alpha_i | \phi)$ $O_{i} (U, α_{i} ∣ ϕ)$ holds if and only if:
  1. $\phi$ holds after the execution of action $\alpha_i$ (dynamic condition).
  2. The action-generated submodel resulting from $\alpha_i$ has the maximal expected deontic value compared to other actions (normative condition).
- This formulation adheres to the STIT principle: the agent "sees to it" that the best expected outcome is obtained.
Axiomatization:
- The paper provides a complete axiomatization for DAML. It includes standard AML axioms, new axioms for expectation atoms (E1, E2), and a set of reduction axioms (R1–R6).
- These reduction axioms allow any formula containing the deontic operator $O_i$ to be translated into an equivalent formula in the language of AML (without deontic operators), facilitating proofs of soundness and completeness.

Key Contributions

Integration of Dynamics and Norms: DAML is the first framework to derive obligations using the machinery of Action Model Logic, explicitly linking the epistemic consequences of actions to their normative evaluation.
Scalar Deontic Semantics: The logic models deontic statements in a scalar way, assigning desirability values to worlds and computing expected values, thereby avoiding the binary "all-or-nothing" pitfalls of traditional deontic logic in uncertain environments.
Formal Handling of Conditional Obligations: The logic introduces a rigorous definition for conditional obligations ( $O_i(\alpha_i | \phi)$ ) that depend on both the outcome of the action and the maximization of expected deontic value.
Proof of Soundness and Completeness: The paper establishes that the DAML proof system is sound and complete with respect to its semantics. This is achieved by reducing the deontic language to the known complete language of AML.

Results and Applications

The paper demonstrates the expressive power of DAML through two primary applications:

The Miners' Puzzle:
- The framework successfully models the scenario where an agent must choose between blocking shaft A, blocking shaft B, or blocking neither.
- By calculating the expected deontic values of the resulting submodels, DAML correctly identifies that "blocking neither shaft" is the obligatory action, as it maximizes the expected number of lives saved given the agent's ignorance of the miners' location. This contrasts with actions that might save all miners if the location were known but risk killing all if guessed incorrectly.
Obligation to Inform (Multi-Agent Scenario):
- A complex scenario involving two agents (Alice and Bethany) and a patient with an allergy is modeled.
- The logic captures the asymmetry of information: Bethany knows the patient is allergic, while Alice does not.
- DAML derives that Bethany has an obligation to inform Alice ( $O_b(\delta_b | \dots)$ ) because, conditional on Alice knowing the allergy, Alice would be obligated to administer the safe drug. The framework successfully reasons through the hypothetical chain of actions and information updates to determine the normative requirement for the informed agent.

Significance and Claims

The paper claims that DAML provides a principled framework for action selection in norm-governed multi-agent systems operating under epistemic uncertainty. Its significance lies in:

Unified Reasoning: It allows agents to evaluate the consequences of alternative actions before execution, considering both the reduction of uncertainty and the normative desirability of outcomes.
Handling Intermediate Cases: The framework supports reasoning in cases where information is incomplete and no single criterion (certainty or outcome) is fully optimized, a common feature in real-world decision-making.
Foundational Contribution: By grounding deontic logic in the dynamic machinery of AML, the paper offers a robust method for formalizing how agents should act when they do not know the full state of the world but must still adhere to normative constraints.

The authors maintain a modest scope, noting that the current work assumes uniform probability over accessible worlds and objective desirability values. Future work is suggested to include explicit probability measures, graded belief operators, and more complex action languages to handle non-deterministic choices and conflicting obligations.