A leading adaptive activation function for deep reinforcement learning

Abstract

The activation function provides deep reinforcement learning with the capability to solve nonlinear problems. However, traditional activation functions have fixed parameter settings and cannot be adjusted adaptively based on constantly changing environmental conditions. This limitation frequently leads to slow convergence speed and inadequate performance of trained agents when confronted with highly complex nonlinear problems. This paper proposes a new method to enhance the ability of reinforcement learning to handle nonlinear problems. This method is mainly divided into two parts. Firstly, an activation function parameter initialization strategy based on environmental characteristics is adopted. Secondly, the Adam algorithm is used to dynamically update the activation function parameters. The activation function proposed in this paper is compared with both traditional activation functions and state-of-the-art activation functions through two experiments. Experimental data show that compared to ReLu, TanH, APA, and EReLu, its convergence speed in DQN tasks is improved by 3.89, 1.29, 0.981, and 2.173 times, respectively, and in SAC tasks, it is improved by 1.504, 1.013, 1.017, and 1.131 times, respectively. The results demonstrate that when the agent utilizes LaTanH as the activation function, it exhibits significant advantages in terms of convergence speed and performance and alleviates the problems of bilateral saturation and gradient vanishing.