Convolutional Neural Network Architecture Seach with Q-Learning

We seeks to automate the process of designing the architecture of the convolutional neural network using reinforcemnet learning. The Qlearning agent is trained to sequentially select the CNN layers to achieve maximum accuracy on the validation set. We limited the layers that the agent can select the convolutional layers, the maximum pooling layers, and the softmax layer with pre-defined hyperparameters such as number of filters and kernel size. We experiment the Q-learning agent on three datasets. First, to test that the agent can converge and select the optimal architecture. We run the Q-learning agent on the randomply generated data from the preselected CNN architecture. The agent is tested on the two standard classification datasets, namely ProstateX and CIFAR-10. The agent is able to find the CNN architecture that achieves better test accuracy than the state-of-the-art model on the ProstateX dataset.