My research sits at the intersection of causality, reinforcement learning, active inference, and computational approaches to consciousness. Find my full profile on Google Scholar.
Journal of Consciousness Studies 33(7), 14–34 · 2026
Uses a deterministic artificial agent that displays hedonic place-preference behaviour to examine what such behaviour can—and cannot—tell us about affective consciousness, its physical basis, and experienced free will.
From Animals to Animats 18 SAB 2026 · Accepted
An active-inference agent directs a fixed interoceptive-precision budget toward the bodily channel it currently believes is most needed. This raises learning-phase survival from 19.9% to 41.4% at the same total budget, while reversing the selector reduces survival to 14.4%.
SSRN preprint 10.2139/ssrn.5392003 · 2025
Introduces Sophisticated Learning, an active-inference algorithm that embeds active parameter learning within multi-step planning. In a seasonal foraging task, it learns faster and survives longer in early trials than Sophisticated Inference and Bayes-adaptive reinforcement-learning baselines.
MSc thesis, University of Cape Town 2023
MSc thesis surveying the current landscape of causal methods applied to reinforcement learning — covering key open problems and promising research directions.
Cooperative AI Workshop, NeurIPS 2021 arXiv:2111.06721
Reviews the state of causal reasoning in multi-agent settings and identifies the key open problems at the frontier of causal MARL research.
arXiv arXiv:2107.01460v1 · 2021
Introduces Mava, a set of MARL components and system abstractions built on top of dm-acme for scalable, distributed training across hardware accelerators. Work done at InstaDeep. I contributed to the original system design and implemented the QMIX algorithm. A revised version of the paper and framework was released in 2023.
Honours thesis, University of Cape Town 2020
Surveys counterfactual reasoning methods through the lens of Pearl's causal hierarchy, focusing on their application to sequential decision-making and RL.
Survey manuscript 2020
Surveys model-based RL approaches that use learned world models for planning, exploration, and sample-efficient learning in complex environments.
Search is generated at build time — run a production build (pnpm build && pnpm preview) to try it.