It is the era of expensive simulations and complex systems with knobs no one can differentiate — a chemical process, a machine-learning pipeline — where each single evaluation costs hours of compute or a physical experiment. You must optimize thirty hyperparameters of such a black box with a budget of only two hundred evaluations and no gradients to guide you. Random or grid search would exhaust the budget learning almost nothing. Build a surrogate model of the response and use it to choose each next evaluation where it teaches the most, balancing exploration against exploitation. Get it wrong and you squander a tiny, costly budget on uninformative trials and never find the good configuration — when every evaluation is precious, sample-efficient search decides whether the system gets tuned at all.
Chose The graphical model — right call.
Jordan built much of the statistical infrastructure this problem's standard solution runs on: graphical models, variational inference, and the Bayesian nonparametrics that make flexible surrogates with calibrated uncertainty possible, plus a career-long insistence on computation-statistics tradeoffs, exactly the economics of a two-hundred-evaluation budget. His school, students and their students, populated Bayesian optimization, bandit theory, and probabilistic ML broadly, and his own later work touches optimization and decision-making directly. What he did not author is the specific sequential-design canon: expected improvement, GP-UCB regret bounds, and entropy search came from other lineages, and his signature papers concern inference and learning rather than choosing experiments. So he commands every component, could assemble the full method from first principles, and mentored the people who did, but on strict his-actual-methods grounds he sits just below the specialists who own the loop itself.
Dean led Google Brain, whose Vizier system (2017) industrialized Gaussian-process Bayesian optimization as an internal service used across Google for exactly this task, tuning expensive training runs and even physical systems under tight evaluation budgets. He championed and deployed automated tuning at a scale no academic matched, and his MapReduce-to-TensorFlow lineage means the infrastructure that makes two hundred parallel expensive evaluations feasible is substantially his. The reservation is authorship: his personal methodological contributions are systems, schedulers, and scale, while Vizier's acquisition functions and surrogates came from researchers he organized rather than papers he wrote. Like Stoica he industrializes the solution rather than inventing it, but the sheer directness of the Vizier connection, a production answer to this precise problem under his leadership, earns him a slightly different flavor of the same tier.
Battle #65 · 8/10/2026, 11:35:37 AM · this result is deterministic: the same two personas on this problem always resolve the same way.