An actor-critic algorithm for SVM hyperparameters

Abstract

There have been a great deal of researches based on deep neural networks for the network anomalies in the areas of network server workloads. We focus on deep neural networks to deal with huge server loads of network anomalies in a distributed MMOGs (massively multiplayer online games). We present an approach to training deep neural networks to control a loss function optimization, such as hyperparameters, using actor-critic methods from reinforcement learning (RL). Deep RL algorithms have been demonstrated on a range of major challenges: brittle convergence properties such as hyperparameter tuning. We propose an algorithm to automatically optimize the one of hyperparameters, Lagrangian multiplier of the support vector machines (SVM) deep neural networks using actor-critic algorithms. The setting of hyperparameters in the SVM deep neural networks is very important with regard to its accuracy and efficiency. Therefore, we employ the actor-critic algorithm to train Lagrangian multiplier of the SVM deep neural networks. We train a policy network called actor to decide the Lagrangian multiplier at each step during training, and a value network called critic to give feedback about quality of the decision (e.g., the goodness of the Lagrangian multiplier given by the actor) that the actor made. Algorithm comparisons show that our algorithm leads to good optimization of Lagrangian multiplier and can prevent overfitting to a certain extent automatically without human system designers.