Empirical Evaluation of Activation Functions and Kernel Initializers on Deep Reinforcement Learning

Hyperparameter optimization has a considerable impact on system performance. However many researchers use the default hyperparameter configurations provided by the machine learning frameworks they use because it require a lot of effort, time and computing resources. In this paper, we analyze the effect of the activation function and the kernel initializer on the performance of deep reinforcement learning. In detail, we compared the performance of 42 combinations of 7 activation functions and 6 kernel initializers on two different network types, namely, multi-layer perceptron (MLP) and convolutional neural network (CNN). From the results, it has been revealed that not only did the optimal combinations differ but also, even with the same hyperparameter, the performance difference was dramatic depending on the network type. This phenomenon was prominent in the kernel initializer.