Hyperparameter optimization of neural networks based on Q-learning

Machine learning algorithms are sensitive to hyperparameters, and hyperparameter optimization techniques are often computationally expensive, especially for complex deep neural networks. In this paper, we use Q-learning algorithm to search for good hyperparameter configurations for neural networks, where the learning agent searches for the optimal hyperparameter configuration by continuously updating the Q-table to optimize hyperparameter tuning strategy. We modify the initial states and termination conditions of Q-learning to improve search efficiency. The experimental results on hyperparameter optimization of a convolutional neural network and a bidirectional long short-term memory network show that our method has higher search efficiency compared with tree of Parzen estimators, random search and genetic algorithm and can find out the optimal or near-optimal hyperparameter configuration of neural network models with minimum number of trials.