Reinforcement learning
How agents can learn from interaction and reuse knowledge across related tasks.
Research
How do agents learn
to act in the world?
My work explores learning, abstraction and decision making in autonomous systems, with a growing focus on safety and meaningful interaction with people.
Research questions
Learning that carries forward.
An intelligent agent should be able to use what it has learned before, make decisions under uncertainty and adapt when the world changes.
At the Robotics, Autonomous Intelligence and Learning (RAIL) Lab at Wits, we investigate reinforcement learning, transfer and composition, multi-agent interaction, robotics and AI safety. The work is grounded in technical questions, with consequences for how intelligent systems operate among people.
Themes
Across the lab and beyond.
Decision making under uncertainty
How autonomous systems choose actions when information is incomplete or conditions change.
Safety for autonomous agents
How to evaluate and constrain agents that learn and act over time in environments shared with people.
Science of intelligence
Connecting questions about machine, human and animal intelligence through the Wits MIND Institute.
Publications
Selected work and the full record.
- 2026 · AAMASMoralityGym: evaluating moral alignment in sequential agents ↗
- 2019 · ICMLComposing Value Functions in Reinforcement Learning ↗
- 2017 · ICMLHierarchy Through Composition with Multitask LMDPs ↗
- 2016 · Machine LearningBayesian Policy Reuse ↗
This is a small selection. For recent papers and co-authored work, see the full publication record.
Full publication record At Wits
Research is a collective endeavour.
RAIL brings together students and collaborators working on intelligent agents. MIND creates a broader home for interdisciplinary research into intelligence and its impact on society.
Explore the MIND Institute ↗