The Agent Lab UoM studies learning-based agents that must operate reliably in uncertain environments.

Focus Areas

Reinforcement Learning Under Constraints

• Safe exploration and risk-aware policy learning
• Stable policy optimization under deployment limits
• Distributional and uncertainty-aware value estimation

Demonstration and Preference Learning

• Offline and hybrid imitation learning
• Preference-driven policy improvement
• Representation transfer from expert behavior

Multi-Agent Coordination

• Decentralized training for cooperative systems
• Equilibrium-aware strategy learning
• Credit assignment in partially observable teams

Methodology Principles

• Reproducibility: complete experiment recipes and logs
• Transparency: ablations and failure analysis included by default
• Deployment relevance: benchmarks selected for real operating constraints

Infrastructure

Our lab maintains shared training pipelines, reproducible compute environments, and open-source repositories for benchmarking.