Surrogate model-based algorithms for expensive black-box optimization

Speaker: Juli Mueller
U.S. National Renewable Energy Laboratory
Summary:
Computationally expensive black-box optimization tasks arise in a wide variety of applications, including the calibration of simulation models against observation data in climate, combustion, and cosmology, as well as in various design and scheduling tasks to name a few. These problems are characterised by the fact that analytic descriptions of the objective and constraint functions are not available (black box), evaluations of the objective and constraint functions are computationally extremely expensive (hours to days), and derivative information is not accessible and cannot be approximated without invoking a prohibitively large number of function calls. To address these challenges, we use surrogate models and active learning strategies. Here, the surrogate models are computationally inexpensive approximations of the costly black-box functions, and they allow us to optimally select new points in the search space. The surrogate models are updated each time new function values have been obtained. We will discuss solution approaches for problems with various characteristics and demonstrate their efficacy on a variety of applications.
Biography:
Juliane “Juli” Mueller is the manager of the Artificial Intelligence, Learning, and Intelligent Systems (ALIS) group within the Computational Science Center at NREL. Juli’s background is in the development of numerical optimization algorithms for black-box and compute-intensive problems where analytic descriptions of objective and constraint functions are not available. Her algorithm developments include surrogate modeling and active learning. In the past, she has applied these optimization algorithms to a variety of U.S. Department of Energy -relevant problems, including environmental applications, fuel search, quantum computing, and high-energy physics. Most recently, Juli’s work as focused on tuning deep learning model architectures with the goal to find models that make robust and reliable predictions. As group leader of ALIS, it is her goal to develop optimization and machine learning capabilities that enable researchers across all NREL applications to accelerate their science.