Robust Autonomous Adaptive Experimentation

Four Specific Aims were proposed to develop, implement, and empirically validate Bayesian learning algorithms for autonomous adaptive experimentation in cognitive science and materials science. Aims 1 and 2, which focused on algorithm development, were achieved by implementing a robust autonomous adaptive system (RAAS, i.e., an experimentation framework) using three algorithms: (1) adaptive design optimization (ADO, a model-based algorithm, Aim 1), (2) Bayesian optimization (BO, model-free algorithm, Aim 2), and (3) Gaussian Process Active Learning (GPAL), a second model-free approach to optimal experimental design that is solely data-driven. Aim 3 tested ADO and GPAL in the fields of decision making and numerical estimation. Aim 4 tested the use of BO to improve the growth of carbon nanotubes and improve the precision of 3D printing (Aim 4 was carried out in collaboration with Dr. Benji Maruyama of the Materials and Manufacturing Directorate at AFRL). In all application domains, we have successfully demonstrated the robustness and efficiency of these algorithms in achieving the research objectives. This work advances the current state of the art in autonomous research in the cognitive and materials sciences.