Python Reinforcement Learning Projects

Generating and training child networks

Research on algorithms that generate neural architectures has been around since the 1970's. What sets NAS apart from previous works is its ability to cater to large-scale deep learning algorithms and its formulation of the task as a reinforcement learning problem. More specifically, the agent, which we will refer to as the Controller, is a recurrent neural network that generates a sequence of values. You can think of these values as a sort of genetic code of the child network that defines its architecture; it sets the sizes of each convolutional kernel, the length of each kernel, the number of filters in each layer, and so on. In more advanced frameworks, the values also determine the connections between ...