Incorporating Human Expertise into Bayesian Optimization

Human expertise presents a grand challenge in real-world design tasks from drug discovery, materials engineering, app design via A/B testing and other domains with time-consuming-to-evaluate objective functions. On the one hand, human experts possess knowledge critical to finding good designs with a small number of objective function evaluations. On the other hand, it is difficult to express this knowledge in a language current algorithms can understand.  For example:

• Human experts that design drugs use knowledge of chemistry and biology to relate easier-to-measure fundamental properties of a candidate drug, like binding affinity and ability to cross cell membranes, to objective functions like efficacy and safety. 

• Product managers at internet companies use their expertise to decide which hyperparameters to use in a product launch based on A/B measurements of scores of metrics, but they have difficulty if asked to specify, before seeing results, an objective function or a small high-priority region of the 20+ dimensional Pareto frontier over metrics.

We describe solutions to this problem leveraging an emerging class of Bayesian optimization methods that treat the objective function (efficacy and safety in drug design; the product manager's utility in A/B-testing-based optimization of mobile apps) not as a black box, as is traditionally done, but as a "grey box" into which we have some visibility. These grey-box Bayesian optimization methods model the objective function computation as a network of functions that pass inputs and outputs to each other. We describe how these methods are being used to help scientists design proteins that control the formation of water ice, with applications in organ preservation for transplantation, and to help product managers at a major internet company improve their products.

Bio: Peter Frazier is the Eleanor and Howard Morgan Professor of Operations Research and Information Engineering at Cornell University. He is also a Senior Staff Scientist at Uber. His research is at the interface between machine learning and operations research including Bayesian optimization and multi-armed bandits. During the pandemic, he led Cornell's COVID-19 Mathematical Modeling Team, which helped design Cornell's asymptomatic COVID-19 testing program and provided university leadership with science-based decision support. At Uber, he managed UberPool's data science group and currently helps to design Uber's pricing systems. He is the winner of best paper awards from the ACM Conference on Economics and Computation, the INFORMS Applied Probability Society, the INFORMS Computing Society, and Winter Simulation Conference.

Integrating Data-and Knowledge-Driven Approaches to Automated Scientific Modelling

In knowledge-driven modelling, an expert derives a model based on their knowledge of the domain studied: Both the structure and the parameters of the model are derived by the expert from knowledge about the entities and processes in the modelled system. In data-driven modelling, many model structures are considered in a trial-and-error fashion, their parameters are fit to data, and a complete model is returned: This is typically a black-box process that does not take into account domain knowledge. Explainable scientific models need to be expressed in formalisms accessible to humans and learned through approaches that integrate data-driven and knowledge-driven modeling and use both data and domain knowledge. 

The talk will discuss approaches to integrating data-driven and knowledge-driven construction of scientific models. Different formalisms for representing models and domain knowledge will be discussed, including process-based models and context-free grammars. The talk will conclude with a discussion of recent approaches that rely on the use of probabilistic context-free grammars and other generative models for equation discovery. 

Bio: Sašo Džeroski is head of the Department of Knowledge Technologies at the Jožef Stefan Institute (Ljubljana, Slovenia). His department develops artificial intelligence methods for machine learning and decision support and uses them to solve practical problems from agriculture and environmental sciences, medicine and life sciences, and space operations/ Earth observation. His own research focuses on artificial intelligence for science, including machine learning from complex data and in the presence of domain knowledge, learning models of dynamical systems, and developing ontologies for describing different branches of computer science research. 

Sašo Džeroski is currently the vicepresident of the Slovenian Artificial Intelligence Society. He has the title of a fellow of EurAI, the European Association of Artificial Intelligence, awarded in recognition for his “Pioneering Work in AI and Outstanding Service for the European AI community”. He is a member of the Macedonian Academy of Sciences and Arts and a member of Academia Europea (European Academy of Humanities, Letters and Sciences).