No Pizza for You: Value-based Plan Selection in BDI Agents

Stephen Cranefield, Michael Winikoff, Virginia Dignum, Frank DignumIJCAI 2017 · DOI

TL;DR

This paper introduces a novel method for BDI (Belief-Desire-Intention) agents to select plans based on user-defined values, moving beyond simple utility or resource optimization. This approach allows agents to make more ethically aligned decisions and provide better explanations for their actions, crucial for applications like elder care robots.

What Problem Does This Paper Address?

As AI systems become more integrated into our lives, especially in sensitive areas like elder care, they need to make decisions that align with human values and ethical principles. Current AI planning often relies on static prioritization or optimization of utility, resources, or time. This approach is insufficient when agents interact with users and must consider complex, context-dependent, and potentially conflicting values (e.g., user's desire for unhealthy food vs. their long-term health goals). The paper addresses the challenge of enabling agents to reason about and select plans that are not just efficient but also ethically sound and aligned with user preferences and societal norms.

Methods

The authors propose extending BDI agents with a value-based reasoning mechanism for plan selection. They leverage the Schwartz theory of basic human values, which posits a universal set of values with individual priority differences.

Key aspects of their method include:

  • Value Hierarchy: Abstract user values (e.g., Self-Enhancement, Conservation) are translated into concrete values (e.g., "follow user's desires," "ensure user's health") and then linked to specific system goals.
  • Goal-Plan Trees with Value Annotations: The standard goal-plan tree representation used in BDI agents is extended to include annotations that specify how each plan or action affects the agent's "Value State" (e.g., increases or decreases satisfaction of a particular value).
  • Value State and Target: Agents maintain a "Value State" (S(v)) representing the current level of satisfaction for each value and a "Target" (T(v)) representing the desired level.
  • Current Importance (CI): The importance of a value in a given situation is calculated based on its salience (how relevant it is to the current goal) and the difference between its target and current state.
  • Constraint Solving: A constraint satisfaction problem is formulated from the annotated goal-plan tree. An external constraint solver is used to find the "best" course of action by minimizing a weighted sum of the Current Importance of values. This approach allows for balancing conflicting values.
  • AgentSpeak Extension: The BDI agent language AgentSpeak (specifically, the Jason implementation) is extended with a VBR(G) construct to trigger value-based reasoning for sub-goal G. This construct is pre-processed into constraints for an external solver.

Datasets

Not discussed in the paper. The paper uses a simulated scenario for illustration and evaluation.

Key Insights and Ideas

  • Values as Stable Decision Drivers: Human values are relatively stable over a lifespan and provide a consistent underlying mechanism for decision-making, offering a more robust basis for agent behavior than ad-hoc rules.
  • Balancing Conflicting Values: The model explicitly handles situations where values pull in opposite directions (e.g., desire for pizza vs. health goals) by using a weighted sum of value importance, allowing for nuanced trade-offs.
  • Explainability through Values: By grounding decisions in underlying values, agents can provide simpler and more coherent explanations for their actions, fostering user trust.
  • Context-Dependent Value Importance: The importance of a value is not static but dynamically calculated based on the current situation (salience) and the gap between the desired and actual state of that value.

Key Results

The paper demonstrates the approach using a scenario of an Elderly Care Artificial System (ECAS) tasked with ordering and serving a meal.

  • In a situation with equal salience for desire, health, wealth, and sustainability, the agent chose to get and eat pizza from a local provider.
  • When priorities shifted, the agent could select a frozen meal (if health was prioritized) or even toast (if wealth was the overriding criterion), showcasing the flexibility of the value-based approach.
  • The computational mechanism was implemented and tested using MATLAB with the YALMIP optimization library and the MOSEK solver, demonstrating that the problem could be solved efficiently (in fractions of a second for the example).

Limitations and Future Work

  • Scalability and Broader Applicability: The authors acknowledge the need to evaluate the scalability of the approach to more complex scenarios and its applicability to a wider range of agent systems.
  • Explanation Generation: While the approach facilitates explainability, a dedicated mechanism for generating natural language explanations of an agent's value-based decisions is a key area for future work.
  • Multi-Agent Decision Making: Extending the framework to handle value-based decision-making in multi-agent systems is another important direction.
  • Decay Pattern Specification: The precise decay pattern for Value States needs further specification.
  • Query Goals: The current implementation has limitations in handling more complex uses of query goals.

Summary updated

BibTeX
@inproceedings{cranefield2017no,
  author = {Stephen Cranefield and Michael Winikoff and Virginia Dignum and Frank Dignum},
  title = {{No Pizza for You: Value-based Plan Selection in BDI Agents}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2017},
  doi = {10.24963/ijcai.2017/26},
}
Figure