Training algorithm

The training steps for the Controller is as follows:

  • For each episode, do the following:
    1. Generate m child network architectures
    2. Train child networks on given task and obtain m validation accuracies
    3. Calculate 
    4. Update 

In Zoph et al., the training procedure is done with several copies of the Controller. Each Controller is parameterized by , which itself is stored in a distributed manner among multiple servers, which we call ...

Get Python Reinforcement Learning Projects now with O’Reilly online learning.

O’Reilly members experience live online training, plus books, videos, and digital content from 200+ publishers.