Public defence in Computer Science, M.Sc., M.Soc.Sc., Petrus Mikkola

Title of the thesis: Humans as Information Sources in Bayesian Optimization

Doctoral student: Petrus Mikkola
Opponent: Professor David Leslie, Lancaster University, United Kingdom
Custos: Professor Samuel Kaski, Aalto University School of Science, Department of Computer Science

There is a growing consensus that artificial intelligence (AI) and machine learning (ML) solutions do not replace humans in the job market but rather complement human strengths and weaknesses, and transform the ways jobs are done. Humans are likely to continue to play a crucial role in any AI system deployed to address challenges in science and industry. Yet, there are many open questions, such as how to access and utilize the knowledge of individual domain experts. In this thesis, the question of how to model human knowledge, beliefs, and preferences is studied in a general context of expert knowledge elicitation, and then in the context of optimization problems. Examples of the latter include "finding the most stable molecule structure used in a solar cell" or "determining which battery charging protocol best balances fast charging with battery degradation." 

The first perspective is that interacting with humans should focus on querying relative information (e.g., comparing two hypotheses A and B) rather than absolute information (e.g., providing a numeric value for hypothesis A). The thesis proposes an ML algorithm that interactively generates preferential questions that require the expert to compare alternatives. Based on the feedback received, it learns a model reflecting the expert's beliefs. This model can be utilized in an ML system to leverage expert knowledge. For example, it allows one to identify which molecular structures are promising candidates for further optimization. 

The second perspective delves deeper into the question of how the elicited expert knowledge can be integrated into optimization algorithms. Viewing the expert as an auxiliary information source in so-called multi-fidelity Bayesian optimization enables selecting informative question to pose to the expert and then integrating this knowledge with other sources of information, such as scientific experiments. 

Every task has its domain experts, and integrating their knowledge into ML systems is often possible, provided that the method of querying them is carefully considered. Using preferential queries is one approach to make the elicitation problem tractable from an ML perspective, while still framing the questions in a format that allows humans to provide meaningful responses.

Key words: machine learning, expert knowledge, elicitation, Bayesian optimization

Thesis available for public display 10 days prior to the defence at: https://aaltodoc.aalto.fi/doc_public/eonly/riiputus/

Contact information:

Doctoral thesis in the School of Science: https://aaltodoc.aalto.fi/handle/123456789/52