Joint-Embedding Predictive Architecture (JEPA) is increasingly used for visual representation learning and as a component in model-based RL, but its behavior remains poorly understood. We provide a theoretical characterization of a simple, practical JEPA variant that has an auxiliary regression head trained jointly with latent dynamics. We prove a No Unhealthy Representation Collapse theorem: in deterministic MDPs, if training drives both the latent-transition consistency loss and the auxiliary regression loss to zero, then any pair of non-equivalent observations, i.e., those that do not have the same transition dynamics or auxiliary value, must map to distinct latent representations. Thus, the auxiliary task anchors which distinctions the representation must preserve. Controlled ablations in a counting environment corroborate the theory and show that training the JEPA model jointly with the auxiliary head generates a richer representation than training them separately. Our work indicates a path to improve JEPA encoders: training them with an auxiliary function that, together with the transition dynamics, encodes the right equivalence relations.

Abstract--We propose MoRe-ERL, a framework that combines Episodic Reinforcement Learning (ERL) and residual learning, which refines preplanned reference trajectories into safe, feasible, and efficient task-specific trajectories. This framework is general enough to incorporate into arbitrary ERL methods and motion generators seamlessly. MoRe-ERL identifies trajectory segments requiring modification while preserving critical task-related maneuvers. Then it generates smooth residual adjustments using B-Spline-based movement primitives to ensure adaptability to dynamic task contexts and smoothness in trajectory refinement. Experimental results demonstrate that residual learning significantly outperforms training from scratch using ERL methods, achieving superior sample efficiency and task performance. Hardware evaluations further validate the framework, showing that policies trained in simulation can be directly deployed in real-world systems, exhibiting a minimal sim-to-real gap. OBOTIC applications, such as multi-arm cooperation, often require frequent motion adaptation to ensure safety and task efficiency.

The application of Reinforcement Learning (RL) to economic modeling reveals a fundamental conflict between the assumptions of equilibrium theory and the emergent behavior of learning agents. While canonical economic models assume atomistic agents act as `takers' of aggregate market conditions, a naive single-agent RL simulation incentivizes the agent to become a `manipulator' of its environment. This paper first demonstrates this discrepancy within a search-and-matching model with concave production, showing that a standard RL agent learns a non-equilibrium, monopsonistic policy. Additionally, we identify a parametric bias arising from the mismatch between economic discounting and RL's treatment of intertemporal costs. To address both issues, we propose a calibrated Mean-Field Reinforcement Learning framework that embeds a representative agent in a fixed macroeconomic field and adjusts the cost function to reflect economic opportunity costs. Our iterative algorithm converges to a self-consistent fixed point where the agent's policy aligns with the competitive equilibrium. This approach provides a tractable and theoretically sound methodology for modeling learning agents in economic systems within the broader domain of computational social science.

N-agent ad hoc teamwork (NAHT) is a newly introduced challenge in multi-agent reinforcement learning, where controlled subteams of varying sizes must dynamically collaborate with varying numbers and types of unknown teammates without pre-coordination. The existing learning algorithm (POAM) considers only independent learning for its flexibility in dealing with a changing number of agents. However, independent learning fails to fully capture the inter-agent dynamics essential for effective collaboration. Based on our observation that transformers deal effectively with sequences with varying lengths and have been shown to be highly effective for a variety of machine learning problems, this work introduces a centralized, transformer-based method for N-agent ad hoc teamwork. Our proposed approach incorporates historical observations and actions of all controlled agents, enabling optimal responses to diverse and unseen teammates in partially observable environments. Empirical evaluation on a StarCraft II task demonstrates that MAT-NAHT outperforms POAM, achieving superior sample efficiency and generalization, without auxiliary agent-modeling objectives. Keywords: multi-agent reinforcement learning, ad hoc teamwork, transformers, agent modeling Acknowledgements This work has taken place in the Learning Agents Research Group (LARG) at UT Austin.

Time-inhomogeneous finite-horizon Markov decision processes (MDP) are frequently employed to model decision-making in dynamic treatment regimes and other statistical reinforcement learning (RL) scenarios. These fields, especially healthcare and business, often face challenges such as high-dimensional state spaces and time-inhomogeneity of the MDP process, compounded by insufficient sample availability which complicates informed decision-making. To overcome these challenges, we investigate knowledge transfer within time-inhomogeneous finite-horizon MDP by leveraging data from both a target RL task and several related source tasks. We have developed transfer learning (TL) algorithms that are adaptable for both batch and online $Q$-learning, integrating valuable insights from offline source studies. The proposed transfer $Q$-learning algorithm contains a novel {\em re-targeting} step that enables {\em cross-stage transfer} along multiple stages in an RL task, besides the usual {\em cross-task transfer} for supervised learning. We establish the first theoretical justifications of TL in RL tasks by showing a faster rate of convergence of the $Q^*$-function estimation in the offline RL transfer, and a lower regret bound in the offline-to-online RL transfer under stage-wise reward similarity and mild design similarity across tasks. Empirical evidence from both synthetic and real datasets is presented to evaluate the proposed algorithm and support our theoretical results.

Sim2Real aims at training policies in high-fidelity simulation environments and effectively transferring them to the real world. Despite the developments of accurate simulators and Sim2Real RL approaches, the policies trained purely in simulation often suffer significant performance drops when deployed in real environments. This drop is referred to as the Sim2Real performance gap. Current Sim2Real RL methods optimize the simulator accuracy and variability as proxies for real-world performance. However, these metrics do not necessarily correlate with the real-world performance of the policy as established theoretically and empirically in the literature. We propose a novel framework to address this issue by directly adapting the simulator parameters based on real-world performance. We frame this problem as a bi-level RL framework: the inner-level RL trains a policy purely in simulation, and the outer-level RL adapts the simulation model and in-sim reward parameters to maximize real-world performance of the in-sim policy. We derive and validate in simple examples the mathematical tools needed to develop bi-level RL algorithms that close the Sim2Real performance gap.

This dissertation makes three main contributions. First, We identify a new connection between policy gradient and dynamic programming in MMDPs and propose the Coordinate Ascent Dynamic Programming (CADP) algorithm to compute a Markov policy that maximizes the discounted return averaged over the uncertain models. CADP adjusts model weights iteratively to guarantee monotone policy improvements to a local maximum. Second, We establish sufficient and necessary conditions for the exponential ERM Bellman operator to be a contraction and prove the existence of stationary deterministic optimal policies for ERM-TRC and EVaR-TRC. We also propose exponential value iteration, policy iteration, and linear programming algorithms for computing optimal stationary policies for ERM-TRC and EVaR-TRC. Third, We propose model-free Q-learning algorithms for computing policies with risk-averse objectives: ERM-TRC and EVaR-TRC. The challenge is that Q-learning ERM Bellman may not be a contraction. Instead, we use the monotonicity of Q-learning ERM Bellman operators to derive a rigorous proof that the ERM-TRC and the EVaR-TRC Q-learning algorithms converge to the optimal risk-averse value functions. The proposed Q-learning algorithms compute the optimal stationary policy for ERM-TRC and EVaR-TRC.

In safety-critical domains such as robotics, navigation and power systems, constrained optimization problems arise where maximizing performance must be carefully balanced with associated constraints. Safe reinforcement learning provides a framework to address these challenges, with Lagrangian methods being a popular choice. However, the effectiveness of Lagrangian methods crucially depends on the choice of the Lagrange multiplier $λ$, which governs the trade-off between return and constraint cost. A common approach is to update the multiplier automatically during training. Although this is standard in practice, there remains limited empirical evidence on the robustness of an automated update and its influence on overall performance. Therefore, we analyze (i) optimality and (ii) stability of Lagrange multipliers in safe reinforcement learning across a range of tasks. We provide $λ$-profiles that give a complete visualization of the trade-off between return and constraint cost of the optimization problem. These profiles show the highly sensitive nature of $λ$ and moreover confirm the lack of general intuition for choosing the optimal value $λ^*$. Our findings additionally show that automated multiplier updates are able to recover and sometimes even exceed the optimal performance found at $λ^*$ due to the vast difference in their learning trajectories. Furthermore, we show that automated multiplier updates exhibit oscillatory behavior during training, which can be mitigated through PID-controlled updates. However, this method requires careful tuning to achieve consistently better performance across tasks. This highlights the need for further research on stabilizing Lagrangian methods in safe reinforcement learning. The code used to reproduce our results can be found at https://github.com/lindsayspoor/Lagrangian_SafeRL.

Traditional PID controllers have limited adaptability for plasma shape control, and task-specific reinforcement learning (RL) methods suffer from limited generalization and the need for repetitive retraining. To overcome these challenges, this paper proposes a novel framework for developing a versatile, zero-shot control policy from a large-scale offline dataset of historical PID-controlled discharges. Our approach synergistically combines Generative Adversarial Imitation Learning (GAIL) with Hilbert space representation learning to achieve dual objectives: mimicking the stable operational style of the PID data and constructing a geometrically structured latent space for efficient, goal-directed control. The resulting foundation policy can be deployed for diverse trajectory tracking tasks in a zero-shot manner without any task-specific fine-tuning. Evaluations on the HL-3 tokamak simulator demonstrate that the policy excels at precisely and stably tracking reference trajectories for key shape parameters across a range of plasma scenarios. This work presents a viable pathway toward developing highly flexible and data-efficient intelligent control systems for future fusion reactors.

Although there is an extensive body of work characterizing the sample complexity of discounted-return offline RL with function approximations, prior work on the average-reward setting has received significantly less attention, and existing approaches rely on restrictive assumptions, such as ergodicity or linearity of the MDP. In this work, we establish the first sample complexity results for average-reward offline RL with function approximation for weakly communicating MDPs, a much milder assumption. To this end, we introduce Anchored Fitted Q-Iteration, which combines the standard Fitted Q-Iteration with an anchor mechanism. We show that the anchor, which can be interpreted as a form of weight decay, is crucial for enabling finite-time analysis in the average-reward setting. We also extend our finite-time analysis to the setup where the dataset is generated from a single-trajectory rather than IID transitions, again leveraging the anchor mechanism.