Generative models for black-box optimization of complex objectives
Event details
| Date | 09.07.2026 |
| Hour | 10:00 › 12:00 |
| Speaker | Edouard Dufour |
| Location | |
| Category | Conferences - Seminars |
EDIC candidacy exam
Exam president: Prof. Nicolas Flammarion
Thesis advisor: Prof. Pascal Fua
Co-examiner: Prof. Alexander Mathis
Abstract
Many high-impact problems in science and engineering reduce to optimizing a complex objective, where each evaluation is costly and often unreliable: physical experiments fail, simulators diverge, and stochastic systems return noisy feedback. Classical black-box optimizers struggle in this regime, where evaluation budgets are tight and the feasible region is hard to characterize a priori. Our work investigates how the statistical expressivity of modern generative samplers can be leveraged in black-box optimization. We propose that these samplers enable efficient navigation of feasible sets and optimization of geometrically complex objectives under limited, unreliable observations.
Two complementary lines of work have already been studied. The main one, SPARROW, demonstrated that a sampler-driven optimizer can substantially reduce the number of function calls needed to maximize complex and unreliable objectives. A parallel line established that samplers can be steered to satisfy hard constraints without sacrificing the statistical diversity of the generated samples. Because SPARROW imposes minimal assumptions on the underlying sampler, these constrained samplers can drive it directly, opening a path to constrained black-box optimization.
Building on these results, two directions are planned. The first is theoretical: a deeper mathematical study of sampler-driven optimization, drawing on the broader literature on sample-efficient black-box optimization, in order to derive stronger guarantees and more efficient algorithms. The second is applied: applying the methods to concrete domains and exploiting the structure those domains expose to sharpen performance beyond what a pure black-box treatment can achieve.
The applications of low-budget black-box optimization are diverse, including engineering design (e.g. aerodynamic shape optimization, mechanical components), scientific discovery (e.g. protein and molecule design under synthesizability and stability constraints), and the safety and security of opaque ML systems (e.g. query-efficient adversarial probing, red-teaming of black-box models).
Selected papers
Exam president: Prof. Nicolas Flammarion
Thesis advisor: Prof. Pascal Fua
Co-examiner: Prof. Alexander Mathis
Abstract
Many high-impact problems in science and engineering reduce to optimizing a complex objective, where each evaluation is costly and often unreliable: physical experiments fail, simulators diverge, and stochastic systems return noisy feedback. Classical black-box optimizers struggle in this regime, where evaluation budgets are tight and the feasible region is hard to characterize a priori. Our work investigates how the statistical expressivity of modern generative samplers can be leveraged in black-box optimization. We propose that these samplers enable efficient navigation of feasible sets and optimization of geometrically complex objectives under limited, unreliable observations.
Two complementary lines of work have already been studied. The main one, SPARROW, demonstrated that a sampler-driven optimizer can substantially reduce the number of function calls needed to maximize complex and unreliable objectives. A parallel line established that samplers can be steered to satisfy hard constraints without sacrificing the statistical diversity of the generated samples. Because SPARROW imposes minimal assumptions on the underlying sampler, these constrained samplers can drive it directly, opening a path to constrained black-box optimization.
Building on these results, two directions are planned. The first is theoretical: a deeper mathematical study of sampler-driven optimization, drawing on the broader literature on sample-efficient black-box optimization, in order to derive stronger guarantees and more efficient algorithms. The second is applied: applying the methods to concrete domains and exploiting the structure those domains expose to sharpen performance beyond what a pure black-box treatment can achieve.
The applications of low-budget black-box optimization are diverse, including engineering design (e.g. aerodynamic shape optimization, mechanical components), scientific discovery (e.g. protein and molecule design under synthesizability and stability constraints), and the safety and security of opaque ML systems (e.g. query-efficient adversarial probing, red-teaming of black-box models).
Selected papers
Practical information
- General public
- Free