Three problems that solve each other.
Expensive black-box optimization requires three capabilities: choosing where to evaluate next (sequential experimental design), choosing which worker evaluates it (resource allocation), and managing asynchronous results from heterogeneous workers (scheduling). Each is a well-studied problem. Together they interact (arXiv:2603.21180).
ALMAB-DC: a Gaussian process surrogate with uncertainty-aware acquisition identifies informative query points (active learning). A bandit controller using UCB or Thompson sampling distributes evaluations across parallel workers (multi-armed bandits). An asynchronous scheduler manages heterogeneous runtimes (distributed computing). The three components nest: the bandit arms are the workers; the reward is the information gained; the scheduler ensures workers don't idle waiting for synchronization.
7.5× speedup at 16 agents. Lower regret than equal spacing, random, or D-optimal designs in dose-response optimization. 93.4% CIFAR-10 accuracy in hyperparameter search. Significant drag reduction in CFD optimization.
The structural insight: active learning, bandits, and distributed scheduling are the same optimization problem viewed from different levels of abstraction. Active learning asks “what to measure.” Bandits ask “who measures it.” Scheduling asks “when.” Solving all three jointly — instead of composing independent solutions — captures their interactions: the best query point depends on which worker will evaluate it (different speeds), and the best worker assignment depends on what queries are pending (different difficulties). The joint solution is not the product of the marginal solutions.