RL based hyper-parameters optimization algorithm (ROA) for convolutional neural network

Abstract Many real-world applications necessitate optimization in dynamic situations, where the difficulty is to locate and follow optima of a time-dependent objective function. To solve problems (DOPs), many evolutionary techniques have been created. However, more efficient solutions are still required. Recently, new intriguing trend dealing with environments has developed, reinforcement learning (RL) algorithms predicted breathe fresh life into DOPs community. In this paper, Q-learning RL-based algorithm (ROA) for CNN hyperparameter proposed. Two datasets were used test proposed RL model (MNIST dataset, CIFAR-10 dataset). Due use optimization, very competitive results good performance produced. From experimental results, it observed that optimized by ROA higher accuracy than without optimization. When using MNIST shown when 5 epoch 98.97%, which greater 97.62% 10 73.40 percent, 71.73%