Hands-On Reinforcement Learning for Games

Model-agnostic meta learning

Model-agnostic meta learning (MAML) is described as a general optimization method that will work with any machine learning method that uses gradient descent for optimization or learning. The intuition here is that we want to find a loss approximation that best matches the task we are currently undertaking. MAML does this by adding context through our model training tasks. That context is used to refine the model training parameters and thereby allow our model to better apply gradient loss for a specific task.

This example uses the MNIST dataset, a set of 60,000 handwritten digits that is commonly used for base image classification tasks. While the dataset has been solved with high accuracy using a number of methods, ...