Exploring the Use of Reinforcement Learning for Selecting Parameters of Numerical Methods in Acoustics
Résumé
Numerical simulation is a important item in the toolbox of most if not all acoustic engineers and researchers. A number of different methods exist, allowing to retrieve time or frequency responses of systems under development. Even though advances have been made, choosing the parameters of the each method to attain optimal resolution within a reasonable time might take some effort for inexperienced but also expert users. The choice sometimes can be guided by previous experience but the parameters are often refined via trial and error when the problem to solve is new to the user. Machine learning on the other hand is a vast subject aiming at letting machines autonomously take sensible decision on different questions. Parts of this field are dedicated to unsupervised learning, ie. the machine learns without the help of an expert dataset, and recent results have shown that this approach allowed to explore completely new strategies for well known problems (AlphaGo, Deepmind, etc...). In this contribution, the authors explore the possibilities of reinforcement learning to select and adapt methods parameters in the context of acoustic simulation. This very exploratory contribution aims at providing insights on existing unsupervised machine learning methods and particularly reinforcement learning as well as demonstrating their ability to correctly select mesh and methods parameters to retrieve frequency responses on simple problems.