DescriptionIn this work, we demonstrate how the Livermore Tournament Fast Batch (LTFB) ensemble algorithm is able to efficiently tune hyperparameters and accelerate the time to solution for several cancer drug discovery networks. Drawn from the DOE-NCI Pilot 1 and ECP CANDLE projects we show significantly improved training quality for the "Uno" data set and associated network and a dramatic reduction in the wall-clock time for training the "Combo" network to a fixed level of convergence. LTFB is an ensemble method that creates a set of neural network models and trains each instance of these models in parallel. Periodically, each model selects another model to pair with, exchanges models, and then run a local tournament against held-out tournament datasets. The winning model will continue training on the local training datasets. LTFB is implemented in the Livermore Big Artificial Neural Network toolkit (LBANN), a toolkit optimized for composing multiple levels of parallelism on HPC architectures.