A simulator for reinforcement learning training in the recommendation field

Deep reinforcement learning (DRL) is an unsupervised learning method, which has great commercial value in recommendation scenarios where it is difficult to collect available labeled data. Although very few researchers have begun to do research on the integration of deep reinforcement learning and recommendation methods, the development of these researches is slow because it is difficult to build an online training environment. Therefore, this paper proposes an environment (i.e. user) simulator for training reinforcement learning models in the recommendation field (RL-E Simulator). On the one hand, the simulator builds a user state generation model based on the Generative Adversarial Network (GAN). On the other hand, we propose an rating model based on attention mechanism, and realizes the reward of the simulator to the actions of DRL-based recommendation (i.e. agent). Experimental results show that the simulator provides a low-cost training environment for the DRL-based recommendation.