Deep Q-Networks for Accelerating the Training of Deep Neural Networks

We present a method, codenamed QAN, for improving the generalization ability of a deep neural network (DNN). It achieves this by using a deep Q-network (DQN) to learn policies to accelerate the training of the DNN across episodes. The state features of the DQN are learned from the weight statistics of the DNN during training. The reward function of this DQN is designed to learn policies to minimize the training time needed by that DNN. The actions of the DQN correspond to some optimization choices during training. The code can be downloaded from https://github.com/bigaidream-projects/qan