Explanation and Knowledge Acquisition in Ad Hoc Teamwork
summary
Architecture combines non‑monotonic logical reasoning (action language + answer‑set solving) with on‑the‑fly learning of fast and frugal decision trees to predict teammates and opponents, deployed in Fort Attack and Half Field Offense. It learns learned executability conditions and causal laws from interactions and generates on‑demand relational explanations for action, contrastive, belief and counterfactual queries. Achieves comparable accuracy using about 10,000 examples vs millions needed by deep baselines and improves task performance when learned axioms are added.
AI-written summary · about
chapters (12)
- 0:00Ad hoc teamwork research
- 1:09Team collaboration challenges
- 3:36Reasoning and learning in architecture
- 5:46Knowledge representation
- 9:01Predictive behavior models
- 10:26Explanations for transparency
- 11:28Model accuracy evaluation
- 12:52Architecture performance comparison
- 13:55Axiom learning from observations
- 15:00Explanation examples in Fort Attack
- 17:30Contributions and future work
- 18:35Q&A
related talks
21mAgentic GraphRAG: Autonomous Knowledge Graph Construction and Adaptive Retrieval
22mHOW TO BUILD YOUR OWN AI AGENT: THE COMPLETE ARCHITECTURE FOR ENGINEERING TEAMS
29mHow to Structure Context to Make Your Agents Smarter
12mAgentic GraphRAG: Autonomous Knowledge Graph Construction and Adaptive Retrieval
20mEnhancing RAG with Multi-Agent Integration
31mAgentic AI architecture patterns
16mBuilding Applications with AI Agents
31mDesigning AI Agents with the Model Context Protocol: From Answers to Actions