# Overview

The authors present a population based training (PBT) method for optimizing hyperparameters while training the parameters of the model. The method is based on evolution, and uses “trials”. A trial is a small chunk of training that has a warm-start using parameters from a previous trial.

This is considered “black box hyperparameter optimisation”, because the method does not need to know about the network architecture. The only requirement is that the network can be warm-started by loading weights, and that performance of the network can be measured.

# Method

1. Run $$N$$ trials in parallel, where a trial is $$S$$ steps of training; each trial has different hyperparameters
2. Evaluate the networks produced
3. Select the $$M \lt N$$ networks that will reproduce
4. Mutate the hyperparameters of the parents to produce a total of $$N$$ children (children = trial)