Learning-based Multi-Agent Planning
We are interested in learning-based approaches for coordinating teams of agents in complex environments. Our focus includes combining various machine learning techniques with classical planning methods to achieve scalable and robust multi-agent coordination.
Research Directions
Bayesian Learning for Multi-Agent Systems
We explore Bayesian learning approaches to model uncertainty in multi-agent environments. This includes learning probabilistic models of agent behaviors, environmental dynamics, and interaction patterns to improve planning and coordination.
Reinforcement Learning for Multi-Agent Coordination
Reinforcement learning offers a promising approach for learning coordination policies directly from experience. We are interested in:
- Multi-agent reinforcement learning (MARL): Learning policies for teams of agents
- Combining RL with Large Neighborhood Search: Using learned policies to guide search algorithms
- Imitation learning: Learning from expert demonstrations for lifelong multi-agent path finding
- Policy learning for heterogeneous agent teams: Handling agents with different capabilities
Large Language Models for Multi-Agent Planning
Large Language Models (LLMs) have shown remarkable capabilities in reasoning and planning. We investigate:
- LLM-guided task planning: Using LLMs to decompose complex multi-agent tasks
- Natural language interfaces: Enabling human-AI collaboration in multi-agent scenarios
- Emergent communication protocols: Learning communication strategies through LLM-based agents
Diffusion Models and Flow Matching
Generative models like diffusion models and flow matching offer new perspectives on multi-agent planning:
- Generative path planning: Using diffusion models to generate diverse, high-quality paths
- Flow matching for coordination: Modeling agent flows in continuous spaces
- Combining with classical methods: Integrating generative models with search-based planning
Key Challenges
- Scalability: Applying learning methods to large-scale multi-agent systems
- Generalization: Ensuring learned policies work across different scenarios
- Safety: Guaranteeing safe coordination when using learned policies
- Interpretability: Understanding and explaining learned coordination strategies