TL;DR
This paper introduces a formal mechanism for group decision-making within the SharedPlans framework for collaborative activity. It specifies how groups of agents can dynamically resolve ambiguities in their plans, such as selecting recipes or assigning tasks, by invoking a pre-defined voting-based decision-making process. This work bridges the gap between the abstract notion of shared plans and the concrete steps needed to realize them.
What Problem Does This Paper Address?
The SharedPlans theory (Grosz and Kraus) provides a formal model for how agents collaborate, outlining the mental states (intentions, beliefs, mutual beliefs) required for a group to have a shared plan. However, it leaves open how groups actually make the decisions necessary to flesh out a partial plan into a complete one. For instance, when a group has a general goal but hasn't decided on the specific steps, the recipe, or who does what, SharedPlans doesn't specify the process for resolving these ambiguities. This paper addresses this gap by proposing a concrete, formal mechanism for group decision-making that can be integrated into the SharedPlans framework.
Methods
The core of the paper is a formalization of a generic, voting-based group decision-making mechanism. This mechanism is itself represented as a fixed, fully-specified SharedPlan (FSP). Agents can invoke this mechanism using declarative speech acts. The paper defines the "force" of these speech acts, drawing on Searle's theory of constitutive rules, to ensure that decisions made through this process lead to the establishment of new beliefs and intentions that are consistent with the SharedPlans framework.
Key methodological components include:
- Conditional Actions: Formalized as
if φ then Adopt.Int.To(X)(adopting an intention to do an action) andif φ then Adopt.Int.Th(ψ)(adopting an intention that a proposition hold). - Performatives (Declarations and Assertions): Declarations are used to establish propositions as mutually believed institutional facts, while assertions are used to report beliefs and intentions. The paper defines the conditions under which these performatives are considered "done" by the group.
- Constitutive Rules: Used to define the force of declarative speech acts, akin to rules in games that create the possibility of the game itself. For example, a declaration in a specific context counts as establishing a certain fact.
- Voting Mechanism (Mv): A specific FSP designed for group decision-making. It involves a sequence of speech acts: an initial declaration by one agent to propose a decision, subsequent declarations by other agents to accept or reject, and a final announcement of the result.
- Triggering Conditions: The invocation of the decision-making FSP is triggered by a group's mutual belief in a specific condition (e.g., that a proposal has been made).
Datasets
Not discussed in the paper.
Key Insights and Ideas
- Bridging the Gap: The paper successfully bridges the theoretical gap between abstract SharedPlans and concrete group decision-making processes, providing a formal mechanism for plan elaboration.
- Constitutive Rules for Social Reality: The application of Searle's constitutive rules to define the force of speech acts is a powerful way to formalize how communication creates social facts and commitments within a multi-agent system.
- Generic Decision-Making Mechanism: The voting mechanism is designed to be generic, applicable to various types of decisions (adopting a plan, selecting a recipe, assigning tasks, binding parameters) by adjusting the context and associated commitments.
- Invoking FSPs: The paper formalizes how agents can invoke pre-defined FSPs (like the voting mechanism) based on triggering conditions and mutual beliefs, leading to the adoption of new commitments and plans.
- Dynamic Plan Elaboration: The mechanism directly supports the dynamic expansion of partial SharedPlans, showing how groups can resolve ambiguities and move towards more complete plans.
Key Results
The paper presents several theorems that formally establish the correctness of the proposed mechanism:
- Theorem 1: Shows that if agents mutually believe they hold certain conditional commitments and a triggering condition holds, then the FSP requirements for a method are necessarily met.
- Theorem 2: Demonstrates that a specific sequence of declarations and mutual beliefs related to a voting outcome (group accepted) leads to the group mutually believing that the decision has been made.
- Theorem 3: Proves that if agents use the voting mechanism to decide to adopt a minimal SharedPlan, they will indeed adopt it, establishing the necessary intentions and mutual beliefs.
The paper also outlines how the mechanism can be adapted for other decision types, such as selecting a recipe, by modifying the specific FSP requirements and conditional commitments.
Limitations and Future Work
- Assumed Rationality: The model assumes agents are truthful in their assertions and that they will act according to their intentions. The paper briefly mentions Perrault's work on agents making assertions they don't believe, suggesting this is an area for future exploration.
- Complexity of Commitment: While the paper presents a generic mechanism, the complexity of defining the specific conditional commitments for each type of decision can grow. The authors note that the "general idea" remains the same, but the number of requirements increases with more subacts and agents.
- Weaker Commitment Model: The authors mention using a "weaker model of the group's commitment" for plan elaboration compared to Grosz and Kraus, requiring only mutual belief that agents intend to find a resolution, rather than a stronger commitment to the resolution itself. They suggest this weaker model might be sufficient in many scenarios.
- Future Work: The paper suggests that the techniques provide a solid foundation for formalizing group decision-making based on any FSP-representable mechanism. Further work could explore more complex decision-making scenarios and agent behaviors.
Summary updated