Learning begins with environmental interaction. Reinforcement learning (RL) models this interaction computationally to achieve goals. Standard RL methods thrive on consistent and immediate feedback. Rooted in extreme value theory, we develop methods to find and exploit rare, high-reward “Black Swan” events. We propose a proof-of-concept by teaching an RL agent to speedrun a game. Counterintuitive trajectories often have delayed rewards but lead to superior strategies.

Many problems in scientific computing and computational mathematics are naturally high-dimensional. A core challenge is the fast, accurate and scalable simulation of high-fidelity waves in complex media. In collaboration with Professor Richard Tsai (UT Austin) and Professor Christian Klingenberg (Uni Wuerzburg), we propose an end-to-end framework grounded in partial differential equations [1]. Our architecture enhances an efficient coarse-grid numerical solver with a corrective neural network. The network is trained on high-fidelity data generated by a computationally expensive fine-grid solver. This work also introduces a novel multi-step gradient descent algorithm for efficiently training neural surrogates. The improved stability allows the use of Parareal, a parallel-in-time algorithm to correct high-frequency wave components.