<p>Global optimization without access to gradient information is a central task to many econometric applications because it is an important computational tool to obtain maximum likelihood estimators for very complicated likelihood functions. The estimating gradient descent framework is particularly popular, which uses local functional evaluation to build gradient estimates and performs gradient descent/ascent from multiple initial points. In this work, we study the problem of coordinating multiple estimating gradient descent “threads” in order to pause or terminate unpromising threads early, making the overall computational procedure more efficient. The high-level idea is to make predictions, either conservative or aggressive, on the potential progress of each estimating gradient descent thread in comparison with progress on other threads. Finally, we test our proposed methodology on airline industry data and compare with competitive methods like the genetic algorithm or simulated annealing.</p>