An Analysis of Tabular Reinforcement Learning Algorithms for the Design of Additively Manufactured Acoustic Metamaterials

Abstract

In the modern age of advancing technology, led by developments in artificial intelligence, automation, precision manufacturing, and optimized design, it is now commonplace to see revolutionary technologies arise from the fusion of these fields. When advanced technologies can combine their strengths and integrate symbiotically for a worthwhile application, the results are bound to be transformative. In the case of precision manufacturing and optimized design, there would be remarkable benefits of merging the two through means of artificial intelligence and automation technology. An autonomous manufacturing system, capable of understanding the performance of each part it fabricates and being able to self-optimize the design, would undoubtably alter the way certain parts, or even entire systems, are engineered and devised. In this thesis, we present an analysis of several intelligent design algorithms acting on a computational emulation of a manufacturing testbed. This testbed is a closed-loop, fused deposition modeling system for printing acoustic metamaterials designed to achieve a desired acoustic passband. In theory, by using measured performance feedback gained through experimental fabrications, the intelligent system should be able to learn the optimal regions of the design space and find the design parameters to achieve the passband objective. The several intelligent design methods researched and analyzed in this thesis are that of basic tabular reinforcement learning algorithms. Reinforcement learning is thought to be a promising approach because of its potential to be data-efficient and learn online effectively. Through various hyperparameter studies, we hoped to discover strong performing hyperparameter configurations for each algorithm, and then further understand their performance on a simulated environment of the testbed through trajectory visualization. From these studies, relations of the hyperparameters and the importance of balancing exploration with exploitation were discovered.