Research
My primary research interest is reinforcement learning (RL), particularly enabling RL agents to efficiently adapt to unseen tasks (meta/multi-task RL) by learning "nice" task representations (resp. good coverage policies) in the presence (resp. absence) of training rewards.
As the dynamics of control tasks are commonly governed by physical laws, I embarked upon the quest of developing RL agents that explicitly model dynamics with equations as I believe that it can enable prior knowledge and/or inductive bias injection, sample efficiency gains, better domain randomization (à-la-Sim2Real) and risk-control thanks to interpretability... Practically, this involves using symbolic regression (SR), the search of analytic expressions composed of mathematical operators, e.g. cos, exp, constants and variables. Due to the lack of SR algorithms that infer accurate expressions in reasonable time, I have worked on developing transformer-based models, trained on synthetically-generated datasets, that search with order of magnitudes less time.