Normative Practical Reasoning via Argumentation and Dialogue

Zohreh Shams, Marina De Vos, Nir Oren, Julian PadgetIJCAI 2016 · OpenReview

Abstract

In a normative environment, an agent's actions are directed not only by its goals but also by norms. Potential conflicts among goals and norms make decision-making challenging. We therefore seek to answer the following questions: (i) how should an agent act in a normative environment? and (ii) how can the agent explain why it acted in a certain way? We propose a solution in which a normative planning problem serves as the basis for a practical reasoning approach based on argumentation. The properties of the best plan(s) w.r.t. goal achievement and norm compliance are mapped to arguments that are used to explain why a plan is justified, using a dialogue game.

Introduction

Autonomous agents operating in social or legal contexts—such as delivery robots on public sidewalks or software agents managing corporate budgets—must navigate environments where personal goals often conflict with social rules, or norms. Deciding on a course of action in these scenarios is known as "practical reasoning."

The core challenge lies in two areas: first, defining what constitutes the "best" plan when preferences over goals and rules are unclear; and second, providing a human-understandable explanation for that choice. If an agent decides to violate a norm (e.g., speeding to deliver life-saving medicine), it must be able to defend that decision through structured justification.

Methodology: From Planning to Argumentation

The proposed framework utilizes a multi-layered approach to bridge the gap between automated planning and human explanation. It begins with Normative Planning, an extension of standard STRIPS planning that incorporates facts (fluents), timed actions, goals, and norms (obligations and prohibitions with specific deadlines).

To resolve conflicts, the planning problem is translated into a formal Argumentation Framework (AF). The system defines specific argument schemes for plans, goals, and norms. Interactions between these elements are modeled using Critical Questions (CQs). For example, a "Goal Argument" might attack a "Plan Argument" if the plan fails to satisfy that goal, creating a directed graph of arguments and attacks.

Figure 3: Argumentation frameworks for four distinct plans, illustrating the varying structures of attacks between goals and norms.

When explicit preferences are missing, the framework applies a Democratic Principle to rank plans. This default hierarchy prefers plans that satisfy the maximum number of goals while violating the minimum number of norms, providing a logical tie-breaker for multiple valid paths.

Key Insights: Modeling Conflict and Explanation

A central innovation of this work is using argumentation as a bridge between raw planning data and human-understandable explanation. By treating a plan as a "claim" subject to attack or defense, the agent's internal logic becomes a map for a structured conversation. This is operationalized through a Socratic Dialogue game—a back-and-forth between the agent (Proponent) and a user (Sceptic).

Figure 2: A detailed argumentation framework for a complex plan, showing a network of attacks involving multiple norms and goals.

The agent defends its choices by "winning" the dialogue game, using the "in" arguments from its internal logic to answer "why" questions. For instance, if a user asks why a certain goal was not met, the agent can point to a specific higher-priority norm or conflicting goal that made the alternative less desirable in that context.

Results and Discussion

The framework demonstrates that a Socratic dialogue can effectively "unfold" complex planning logic into a series of simple, defensive statements. Formally, the authors prove that the system satisfies "rationality postulates," ensuring that the conclusions reached are logically consistent and closed. Plans are deemed "justified" using credulous preferred semantics, meaning they are defensible even when other alternatives exist.

While the theoretical foundation is robust, the authors acknowledge that future work is needed to evaluate these explanations with real human users to test their persuasiveness. Additionally, the model could be extended to handle more complex temporal rules and implement preference learning to allow agents to adapt to user values over time.

Summary updated

Abstract

In a normative environment an agent's actions are not only directed by its goals but also by norms. Here, potential conflicts among the agent's goals and norms makes decision-making challenging. We therefore seek to answer the following questions:(i) how should an agent act in a normative environment? and (ii) how can the agent explain why it acted in a certain way? We propose a solution in which a normative planning problem serves as the basis for a practical reasoning approach based on argumentation. The properties of the best plan(s) w.r.t. goal achievement and norm compliance are mapped to arguments that are used to explain why a plan is justified, using a dialogue game.
BibTeX
@inproceedings{shams2016normative,
  author = {Zohreh Shams and Marina De Vos and Nir Oren and Julian Padget},
  title = {{Normative Practical Reasoning via Argumentation and Dialogue}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2016},
  url = {https://openreview.net/forum?id=SkbUNBzuWS},
}
Figure