Course on multi-objective decision making under uncertainty
Date:
A one-day course taught as part of the 2026 SNUTI On Site workshop series, consisting of a 1-hour lecture and a 3-hour hands-on lab run as a team competition. [Event Page]
Lecture [Slides]
Title: AI로 배우는 신약개발 데이터분석 실습: Multi-objective decision making under uncertainty
Overview: Design problems of practical interest often involve multiple objectives. For example, we may want to design antibodies that express well in cell culture, bind strongly to a target antigen, and remain stable at high temperatures. Because these properties can trade off against one another and exhibit complex, heavy-tailed dependence, there is typically no single optimal design. The lecture introduced multi-objective Bayesian optimization (MOBO) as a sample-efficient framework for reasoning about such problems.
1. Multi-objective optimization. Using lead optimization in the clinical-stage antibody landscape as a motivating example, I introduced Pareto dominance, the Pareto set and Pareto front, and the goal of approximating the true Pareto front. I then discussed quality indicators, such as the hypervolume, which score approximate Pareto fronts and enable comparisons across optimization algorithms.
2. Bayesian optimization. After a crash course on the key Bayesian ingredients (conditional probability, predictive distributions, and marginalization), I presented the Bayesian optimization loop: propose candidate designs, predict their properties using a probabilistic surrogate, rank them with an acquisition function that balances exploration and exploitation, evaluate the top candidate, and refit the surrogate model.
3. Multi-objective decision making. The conceptual progression from quality indicator to utility function to acquisition function provides a unified way to understand standard multi-objective acquisition strategies. I presented several commonly used acquisition functions and discussed their respective advantages and limitations. I concluded with two extensions that are important in practice: noisy acquisition functions, which account for uncertainty in the baseline observations, and batch acquisition functions, which select sets of designs for parallel evaluation.
Lab [Starter code]
The 3-hour practicum, implemented in BoTorch, had students work together in antibody design teams. Each round, each team picks a batch of candidate sequences to “test,” aiming to grow the Pareto front and its hypervolume under a fixed experimental budget. The goals were to learn the BoTorch syntax for every concept from the lecture and to build intuition for the trade-offs among acquisition strategies by visualizing the acquired designs. Use of AI coding agents was strongly encouraged for visualization.
The notebooks build up in sequence:
- Notebook 0: warm-up on Pareto fronts and hypervolume
- Notebook 1: a full MOBO loop: seeded, noisy, sequential-greedy batch selection over a pre-generated library of antibody candidates
- Notebooks 2 and 3: practice with the strategy cards and the team competition
- Notebook 4: optional extensions, including information-theoretic acquisition functions
