Distributed deep reinforcement learning for simulation control

Abstract Several applications in the scientific simulation of physical systems can be formulated as control/optimization problems. The computational models for such generally contain hyperparameters, which control solution fidelity and expense. tuning these parameters is non-trivial general approach to manually ‘spot-check’ good combinations. This because optimal hyperparameter configuration search becomes intractable when parameter space large they may vary dynamically. To address this issue, we present a framework based on deep reinforcement learning (RL) train neural network agent that controls model solve by varying First, validate our RL problem controlling chaos chaotic dynamically changing system. Subsequently, illustrate capabilities accelerating convergence steady-state fluid dynamics solver automatically adjusting relaxation factors discretized Navier–Stokes equations during run-time. results indicate run-time learned policy leads significant reduction number iterations compared random selection factors. Our point potential benefits adaptive strategies across different geometries boundary conditions with implications reduced campaign expenses 4 Data codes available at https://github.com/Romit-Maulik/PAR-RL .