Abstract

In the process of the design and correction of room acoustic properties, it is often necessary to select the appropriate type of acoustic treatment devices and make decisions regarding their size, geometry, and location of the devices inside the room under the treatment process. The goal of this doctoral dissertation is to develop and validate a mathematical model that allows predicting the effects of the application of the scattering system in selected points of the room. Further, it is aimed to use this model in the process of computer optimization of the room sound treatment process. The means for achieving these goals are machine learning algorithms with a particular focus on deep learning and reinforcement learning. Deep machine learning models are trained by using computer simulation employing the finite difference method (FDTD), which is used as a source of the reward signal. The simulation model is based on the modified difference equations derived by the author of this thesis that allows simulating the behavior of acoustic diffusers in anechoic conditions. In order to reproduce this type of conditions, the results obtained by the reinforcement learning algorithms, in particular - by the deep policy gradient algorithm, were compared with the results obtained with the classical methods of designing acoustic diffusers and those obtained using another optimization method, i.e., genetic algorithms, where the numerical simulation of the behavior of the acoustic diffuser serves to calculate the value of the fitness function, which plays a role analogous to the reward function. The optimized property of acoustic diffusers is the autocorrelation diffusion coefficient. Numerical experiments have shown that optimization algorithms can be used to maximize the metrics computed by numerical simulation. The best results were obtained with the reinforcement learning algorithms. To validate the calculation results, the measurement of acoustic diffuser prototypes was also performed in the anechoic chamber. As a result of the measurements, the thesis confirmed that the algorithms resulting from computer optimization are characterized by more desirable parameters - the broadband autocorrelation diffusion coefficient and the band diffusion coefficients calculated for the bands with central frequencies with the values from 250 Hz up to 4 kHz. The proposed algorithm is a new approach to the intelligent design of acoustic systems to improve room acoustic properties.

Author (1)