Reinforcement Learning for Adaptive Time-Stepping in the Chaotic Gravitational Three-Body Problem

TL;DR

This paper introduces a reinforcement learning approach to adaptively select time-step sizes in astrophysical simulations, improving accuracy and efficiency without expert intervention, demonstrated on a three-body gravitational problem.

Contribution

It presents a novel reinforcement learning method for dynamic time-step selection that outperforms traditional fixed or expert-tuned approaches in astrophysical simulations.

Findings

Reinforcement learning effectively adapts time steps to system dynamics.

The method reduces computational effort while maintaining accuracy.

It generalizes to other integrators without retraining.

Abstract

Many problems in astrophysics cover multiple orders of magnitude in spatial and temporal scales. While simulating systems that experience rapid changes in these conditions, it is essential to adapt the (time-) step size to capture the behavior of the system during those rapid changes and use a less accurate time step at other, less demanding, moments. We encounter three problems with traditional methods. Firstly, making such changes requires expert knowledge of the astrophysics as well as of the details of the numerical implementation. Secondly, some parameters that determine the time-step size are fixed throughout the simulation, which means that they do not adapt to the rapidly changing conditions of the problem. Lastly, we would like the choice of time-step size to balance accuracy and computation effort. We address these challenges with Reinforcement Learning by training it to select the time-step size dynamically. We use the integration of a system of three equal-mass bodies that move due to their mutual gravity as an example of its application. With our method, the selected integration parameter adapts to the specific requirements of the problem, both in terms of computation time and accuracy while eliminating the expert knowledge needed to set up these simulations. Our method produces results competitive to existing methods and improve the results found with the most commonly-used values of time-step parameter. This method can be applied to other integrators without further retraining. We show that this extrapolation works for variable time-step integrators but does not perform to the desired accuracy for fixed time-step integrators.