Abstract: In recent years, deep learning-based deep belief network (DBN) has achieved successful applications in artificial intelligence and big data prediction analysis. However, too many hidden layers in DBN easily leads to a poor learning accuracy of supervised fine-tuning method (BP algorithm), even failure because of gradient diffusion, and robustness is poor. For this problem, an improved DBN based on reinforcement learning (RL-DBN) is proposed. First, adaptive contrastive divergence (ACD) algorithm is used to fast pre-train the hidden layers of DBN so that the better initial weight can be achieved, then the RL algorithm is used to replace BP algorithm to fine-tune DBN so that higher accuracy and better robustness can be achieved. The experimental results show that, compared with several existing similar models, the proposed RL-DBN achieves better performance in learning rate, accuracy and robustness. 

Key words: Deep belief network, reinforcement learning, adaptive contrastive divergence, robustness

摘要： 近年来，基于深度学习思想发展起来的深度信念网络(Deep Belief Networks, DBN)在人工智能和大数据预测分析中得到了成功的应用。由于DBN的隐含层数较多，传统的DBN有监督精调(Fine-tuning)方法—BP算法很难得到令人满意的学习精度，甚至会因为梯度扩散(Gradient Diffusion)导致精度调节失败，且网络鲁棒性差。针对此问题，提出一种基于强化学习策略的DBN模型(RL-DBN)及其算法。首先利用自适应对比散度(Adaptive Contrastive Divergence, ACD)算法来快速预训练DBN的隐含层以获取较优的初始权值，然后用强化学习算法代替BP算法对DBN进行精调以提高有监督学习的精度和网络的鲁棒性。实验结果表明，相较于现有的类似模型，RL-DBN在学习速度、精度以及鲁棒性能等方面均有较大提高。

关键词: 深度信念网络, 强化学习, 自适应对比散度, 鲁棒性能