I. Introduction
Although deep learning has achieved notable successes and shown its advantages, including efficiency, scalability, and generalization capability, in procedural content generation (PCG), endless online generation of game levels via deep learning remain to be a rarely explored direction [1], [2]. Recently, Shu et al. [3] introduced an experience-driven PCG via reinforcement learning (EDRL) framework, integrated generative adversarial networks (GAN) [4], [5], and deep reinforcement learning (RL) [6] to realize the online level generation (OLG) of game levels while maximizing experience-driven quality measurements. The work in[3] reveals the potential of endless OLG via deep learning.