To be greedy, or not to be — That is the question for Population Based Training variants

Achieving excellent results with neural networks requires careful hyperparameter tuning, which can be automated via hyperparameter optimization algorithms such as Population Based Training (PBT). PBT stands out for its capability to efficiently optimize hyperparameter schedules in parallel and within the wall-clock time of training a single network. Several PBT variants have been proposed that improve performance in the experimental settings considered in the associated publications. However, the experimental settings and tasks vary across publications, while the best previous PBT variant is not always included in the comparisons, thus making the relative performance of PBT variants unclear. In this work, we empirically evaluate five single-objective PBT variants on a set of image classification and reinforcement learning tasks with different setups (such as increasingly large search spaces). We find that the Bayesian Optimization (BO) variants of PBT tend to behave greedier than the non-BO ones, which is beneficial when aggressively pursuing short-term gains improves long-term performance and harmful otherwise. This is a previously overlooked caveat to the reported improvements of the BO PBT variants. Examining their theoretical properties, we find that the returns of BO PBT variants are guaranteed to asymptotically approach the returns of the greedy hyperparameter schedule (rather than the optimal one, as claimed in prior work). Together with our empirical results, this leads us to conclude that there is currently no single best PBT variant capable of outperforming others both when pursuing short-term gains is helpful in the long term, and when it is harmful.