Collaborate and explain on-the-fly: knowledge-based reasoning and learning in ad hoc teamwork

Area of Science:

• Artificial Intelligence
• Multi-agent Systems
• Human-Computer Interaction

Background:

• Current ad hoc teamwork methods rely heavily on large labeled datasets, which are often unavailable in practical scenarios.
• Existing approaches lack transparency and flexibility, hindering adaptation to changing environments or team compositions.
• State-of-the-art methods primarily treat ad hoc teamwork as a learning problem, facing limitations in real-world applicability.

Purpose of the Study:

• To develop a novel architecture for ad hoc teamwork that overcomes the limitations of existing data-driven approaches.
• To enable AI agents to collaborate effectively with others without prior coordination.
• To create a transparent, efficient, and explainable AI system for dynamic teamwork.

Main Methods:

• The architecture combines knowledge-based reasoning (commonsense knowledge) with data-driven methods (rapidly learned behavior models).
• It leverages Large Language Models (LLMs) for predicting agent behavior and understanding abstract goals.
• Non-monotonic logical reasoning is employed for decision-making, integrating prior knowledge and learned models.

Main Results:

• The proposed architecture demonstrates reliable, efficient, transparent, and scalable performance in the VirtualHome simulation environment.
• It significantly outperforms a purely knowledge-based baseline.
• Performance is comparable or superior to purely data-driven baselines, using significantly fewer resources.

Conclusions:

• The integrated approach offers a more practical and robust solution for ad hoc teamwork compared to existing methods.
• The architecture facilitates rapid knowledge revision and provides explainable decision-making capabilities.
• This work advances the field of AI collaboration by enabling adaptable and transparent agent interactions.