Congratulations Dr. Suzan Ece Ada!

Robust and Adaptive Deep Reinforcement Learning  

Abstract  

This thesis addresses the challenge of developing robust and adaptive reinforcement learning (RL) agents that operate in complex, non-Markovian, and non-stationary environments. Unsupervised Meta-Testing with Conditional Neural Processes (UMCNP) addresses few-shot adaptation under unknown dynamics when reward signals are missing at test time. UMCNP learns a dynamics model to enable sample-efficient adaptation through self-generated trajectories. Episodic Return Progress with Bidirectional Progressive Neural Networks (ERP-BPNN) presents a human-inspired framework for multi-task learning by integrating a novel intrinsic motivation signal (ERP) for autonomous task switching with a bidirectional progressive neural network architecture, thereby facilitating effective skill transfer among morphologically different agents. State Reconstruction for Diffusion Policies (SRDP) confronts the challenge of generalization to out-of-distribution states in offline RL by incorporating a state reconstruction loss into the diffusion policy learning process. Forecasting in Non-stationary Offline RL (FORL) mitigates non-trivial non-stationarities by unifying conditional diffusion models with probabilistic zero-shot time-series foundation models. This framework proactively forecasts and corrects for abrupt, hidden observation offsets. Empirical evaluations across a range of continuous control, offline RL benchmarks, and robotics tasks confirm the efficacy of these methods. Our results demonstrate significant improvements in meta-testing sample efficiency, faster convergence via bidirectional skill transfer with return progress, superior generalization to out-of-distribution states, and robust performance against abrupt, non-Markovian shifts in the observation function.