Hands-On Machine Learning for Algorithmic Trading

How adversarial training works

The key innovation of GANs is a new way of learning this probability distribution. The algorithm sets up a competitive—or adversarial—game between two neural networks called the generator and the discriminator, respectively.

The learning objective of the generator is to create output from random input, for example, images of faces. The discriminator, in turn, aims to differentiate between the generator's output and a set of training data that reflects the target output, for example, a database of celebrities as in a popular application. The overall purpose is that both networks get better at their respective tasks while they compete so that the generator ends up producing output that a machine can no longer ...