moritz.txt

In order to adequately explore the large hyperparameter spaces of modern learning models, it is necessary to evaluate a large
number of configurations. Given the growing costs of model training, we would ideally like to perform this search as efficiently
as possible and in parallel. In this session, we will discuss and compare three asynchronous search algorithms that can be
deployed in parallel. Asynchronous algorithms have the advantage that they are robust against stragglers, avoid idle workers
and can be combined with extensive early stopping, making them more efficient. However, these meta-level search algorithms
introduce additional hyperparameters, for example the number of resources allocated to the training of each configuration, 
leading to a trade-off between exploration and exploitation. We want to see how the authors of the three papers design their 
experiments to guarantee comparability of the algorithms.