Proximal policy optimization (PPO) is a deep reinforcement learning algorithm based on the actor–critic (AC) architecture. In the classic AC architecture, the Critic (value) network is used to estimate the value function while the Actor (policy) network optimizes the policy according to the estimated value function. The efficiency of the classic AC architecture is limited due that the policy does not directly participate in the value function update. The classic AC architecture will make the value function estimation inaccurate, which will affect the performance of the PPO algorithm. For improvement, we designed a novel AC architecture with policy feedback (AC-PF) by introducing the policy into the update process of the value function and further proposed the PPO with policy feedback (PPO-PF). For the AC-PF architecture, the policy-based expected (PBE) value function and discount reward formulas are designed by drawing inspiration from expected Sarsa. In order to enhance the sensitivity of the value function to the change of policy and to improve the accuracy of PBE value estimation at the early learning stage, we proposed a policy update method based on the clipped discount factor. Moreover, we specifically defined the loss functions of the policy network and value network to ensure that the policy update of PPO-PF satisfies the unbiased estimation of the trust region. Experiments on Atari games and control tasks show that compared to PPO, PPO-PF has faster convergence speed, higher reward, and smaller variance of reward.

Tumor gene expression data classification via sample expansion-based deep learning Tumor gene expression data classification via sample expansion-based deep learning

In recent years, the proximal policy optimization (PPO) algorithm has received considerable attention because of its excellent performance in many challenging tasks. However, there is still a large space for theoretical explanation of the mechanism of PPO’s horizontal clipping operation, which is a key means to improve the performance of PPO. In addition, while PPO is inspired by the learning theory of trust region policy optimization (TRPO), the theoretical connection between PPO’s clipping operation and TRPO’s trust region constraint has not been well studied. In this article, we first analyze the effect of PPO’s clipping operation on the objective function of conservative policy iteration, and strictly give the theoretical relationship between PPO and TRPO. Then, a novel first-order policy gradient algorithm called authentic boundary PPO (ABPPO) is proposed, which is based on the authentic boundary setting rule. To ensure the difference between the new and old policies is better kept within the clipping range, by borrowing the idea of ABPPO, we proposed two novel improved PPO algorithms called rollback mechanism-based ABPPO (RMABPPO) and penalized point policy difference-based ABPPO (P3DABPPO), which are based on the ideas of rollback clipping and penalized point policy difference, respectively. Experiments on the continuous robotic control tasks implemented in MuJoCo show that our proposed improved PPO algorithms can effectively improve the learning stability and accelerate the learning speed compared with the original PPO.

In recent years, the deep reinforcement learning (DRL) algorithms have been developed rapidly and have achieved excellent performance in many challenging tasks. However, due to the complexity of network structure and a large amount of network parameters, the training of deep network is time-consuming, and consequently, the learning efficiency of DRL is limited. In this paper, aiming to speed up the learning process of DRL agent, we propose a novel approximate policy-based accelerated (APA) algorithm from the viewpoint of the error analysis of approximate policy iteration reinforcement learning algorithms. The proposed APA is proven to be convergent even with a more aggressive learning rate, making the DRL agent have a faster learning speed. Furthermore, to combine the accelerated algorithm with deep Q-network (DQN), Double DQN and deep deterministic policy gradient (DDPG), we proposed three novel DRL algorithms: APA-DQN, APA-Double DQN, and APA-DDPG, which demonstrates the adaptability of the accelerated algorithm with DRL algorithms. We have tested the proposed algorithms on both discrete-action and continuous-action tasks. Their superior performance demonstrates their great potential in the practical applications.

Hyperspectral image (HSI) classification remains a challenging problem due to unique characteristics of HSI data (such as numerous bands and strong correlations in the spectral and spatial domains) and small sample size. To address such concerns, we propose a novel spectral–spatial feature extraction method for HSI classification by employing graph embedding and deep learning (DL) models. Since the conventional graph cannot capture the complex manifold relationship of HSI data, and there exist the observations of within-class variation as well as the similarity between different classes in the spectral domain, we construct the supervised within-class/between-class hypergraph (SWBH) to extract the spectral features of HSI. Since it is difficult for DL models to learn representative features for HSI data when the labeled training samples are limited, we propose the random zero settings to newly generate a large amount of labeled HSI samples for the training of convolutional neural network (CNN). The designed sample expanded CNN (SECNN) is used to extract the HSI spatial features. Thus, the spectral–spatial features of HSI can be learned by integrating the features extracted from SWBH and SECNN, respectively. Experiments on three real HSI datasets demonstrate higher classification accuracy of the proposed SWBH–SECNN method.

For the zero-shot image classification with relative attributes (RAs), the traditional method requires that not only all seen and unseen images obey Gaussian distribution, but also the classifications on testing samples are made by maximum likelihood estimation. We therefore propose a novel zero-shot image classifier called random forest based on relative attribute. First, based on the ordered and unordered pairs of images from the seen classes, the idea of ranking support vector machine is used to learn ranking functions for attributes. Then, according to the relative relationship between seen and unseen classes, the RA ranking-score model per attribute for each unseen image is built, where the appropriate seen classes are automatically selected to participate in the modeling process. In the third step, the random forest classifier is trained based on the RA ranking scores of attributes for all seen and unseen images. Finally, the class labels of testing images can be predicted via the trained RF. Experiments on Outdoor Scene Recognition, Pub Fig, and Shoes data sets show that our proposed method is superior to several state-of-the-art methods in terms of classification capability for zero-shot learning problems.