Aritra Pal

PhD
University of Helsinki (UH)

While Reinforcement Learning (RL) has achieved substantial progress in sequential decision-making, conventional approaches typically rely on the assumption that the underlying environment is stationary and that learned value estimates are sufficiently reliable for deployment. In real-world applications, however, these assumptions are frequently violated; system dynamics, task objectives, observations, and external disturbances often drift or shift over time. Concurrently, an RL agent must navigate decision-making under uncertainty, stemming both from inherent environmental stochasticity (aleatoric) and incomplete knowledge (epistemic). Efficiently unifying online adaptation with rigorous uncertainty estimation remains a critical open problem in modern AI. This research will investigate how an autonomous agent can simultaneously identify changes in environmental conditions, quantify its epistemic and aleatoric uncertainty, and leverage this information to adapt its policy, value estimates, and exploration strategy dynamically. The project will synthesize concepts from meta-RL, continual RL, Bayesian RL, model-based RL, and distributional RL. While existing frameworks, such as VariBAD, FANS-RL, and PEARL demonstrate the efficacy of latent context inference for adaptation, and methods like Bootstrapped DQN, randomized prior functions, and PETS illustrate how uncertainty can guide exploration and model predictive control, these paradigms remain insufficiently integrated. This gap is particularly evident in environments where non-stationarity and high uncertainty occur simultaneously. The expected outcome is a unified algorithmic framework for uncertainty-aware adaptation in RL, supported by theoretical analysis and empirical evaluation. The developed methods will focus on sample-efficient online adaptation, robust transfer learning, and safe exploration, validated across both complex simulated benchmarks and, where feasible, real-world robotic or autonomous systems.

Academic Track
August 1st, 2026 - July 31st, 2030
ELLIS Edge Newsletter
Join the 6,000+ people who get the monthly newsletter filled with the latest news, jobs, events and insights from the ELLIS Network.