模式识别与人工智能
Home      About Journal      Editorial Board      Instructions      Ethics Statement      Contact Us                   中文
Pattern Recognition and Artificial Intelligence
22 Judgement and Disposal of Academic Misconduct Article
22 Copyright Transfer Agreement
22 Proof of Confidentiality
22 Requirements for Electronic Version
More....
22 Chinese Association of Automation
22 National ResearchCenter for Intelligent Computing System
22 Institute of Intelligent Machines,Chinese Academy of Sciences
More....
 
 
2021 Vol.34 Issue.3, Published 2021-03-25
Papers and Reports    Researches and Applications    Research on Reinforcement Learning   
   
Research on Reinforcement Learning
191 Task Allocation Strategy of Spatial Crowdsourcing Based on Deep Reinforcement Learning
NI Zhiwei, LIU Hao, ZHU Xuhui, ZHAO Yang, RAN Jiamin
In the traditional dynamic online task allocation strategy, it is difficult to effectively make use of historical data for learning and the impact of current decisions on future revenue is not taken into account. Therefore, a task allocation strategy of spatial crowdsourcing based on deep reinforcement learning is proposed. Firstly, maximizing long-term cumulative income is regarded as an objective function and the task assignment problem is transformed into the solution of Q value of state action and the one-to-one distribution between workers and tasks by modeling from the perspective of a single crowdsourcing worker grounded on Markov decision process. Secondly, the improved deep reinforcement learning algorithm is applied to learn the historical task data offline to construct the prediction model with respect to Q value. Finally, Q value in real time gained by the model in the dynamic online distribution process is regarded as a side weight of KM algorithm. The optimal distribution of global cumulative returns can be achieved. The results of comparative experiment on the real taxi travel dataset show that the proposed strategy increases the long-term cumulative income while the number of workers is within a certain scale.
2021 Vol. 34 (3): 191-205 [Abstract] ( 1114 ) [HTML 1KB] [ PDF 2559KB] ( 878 )
206 Sequence to Sequence Multi-agent Reinforcement Learning Algorithm
SHI Tengfei, WANG Li, HUANG Zirong
The multi-agent reinforcement learning algorithm is difficult to adapt to dynamically changing environments of agent scale. Aiming at this problem, a sequence to sequence multi-agent reinforcement learning algorithm(SMARL) based on sequential learning and block structure is proposed. The control network of an agent is divided into action network and target network based on deep deterministic policy gradient structure and sequence-to-sequence structure, respectively, and the correlation between algorithm structure and agent scale is removed. Inputs and outputs of the algorithm are also processed to break the correlation between algorithm policy and agent scale. Agents in SMARL can quickly adapt to the new environment, take different roles in task and achieve fast learning. Experiments show that the adaptability, performance and training efficiency of the proposed algorithm are superior to baseline algorithms.
2021 Vol. 34 (3): 206-213 [Abstract] ( 633 ) [HTML 1KB] [ PDF 718KB] ( 858 )
214 Label-Free Network Pruning via Reinforcement Learning
LIU Huidong, DU Fang, YU Zhenhua, SONG Lijuan
To remove redundant structures from deep neural networks and find a network structure with a good balance between capability and complexity, a label-free global compression learning method(LFGCL) is proposed. A global pruning strategy is learned based on the network architecture representation to effectively avoid the appearance of the suboptimal compression rate owing to network pruning in a layer-by-layer manner. LFGCL is independent from data labels during pruning, and the network architecture is optimized by outputting similar features with the baseline network. The deep deterministic policy gradient algorithm is applied to explore the optimal network structure by inferring the compression ratio of all layers through reinforcement learning. Experiments on multiple datasets show that LFGCL generates better performance.
2021 Vol. 34 (3): 214-222 [Abstract] ( 507 ) [HTML 1KB] [ PDF 763KB] ( 420 )
223 Emotion-Based Heterogeneous Multi-agent Reinforcement Learning with Sparse Reward
FANG Baofu, MA Yunting, WANG Zaijun, WANG Hao
In reinforcement learning, the convergence speed and efficiency of the agent are greatly reduced due to its inability to acquire effective experience in an sparse reward distribution environment. Aiming at this kind of sparse reward problem, a method of emotion-based heterogeneous multi-agent reinforcement learning with sparse reward is proposed in this paper. Firstly, the emotion model based on personality is established to provide incentive mechanism for multiple heterogeneous agents as an effective supplement to external rewards. Then, based on this mechanism, a deep deterministic strategy gradient reinforcement learning algorithm based on intrinsic emotional incentive mechanism under sparse rewards is proposed to accelerate the convergence speed of agents. Finally, multi-robot pursuit is used as a simulation experiment platform to construct sparse reward scenarios with different difficulty levels, and the effectiveness and superiority of the proposed method in pursuit success rate and convergence speed are verified.
2021 Vol. 34 (3): 223-231 [Abstract] ( 633 ) [HTML 1KB] [ PDF 1685KB] ( 535 )
Papers and Reports
232 Structure-Preserving Super-Resolution Reconstruction Based on Multi-residual Network
ZHANG Mingjin, PENG Xiaoqi, GUO Jie, LI Yunsong, WANG Nannan, GAO Xinbo
Aiming at the problems of geometric structure distortion and missing details in image super-resolution reconstruction, a structure-preserving super-resolution reconstruction algorithm based on multi-residual network is proposed. Deep learning is carried out in the wavelet transform domain and the gradient domain. Three kinds of residual networks are introduced. The structure and the edge information are reconstructed by the residual gradient network. The high-frequency information of the image is reconstructed as a whole by the residual wavelet transform network. The network attention is adjusted by the residual channel attention network , the important channel features are emphatically learned, and the high frequency information of the image is recovered locally. Experiments show that the proposed algorithm achieves better performance in both quantitative results and visual effects.
2021 Vol. 34 (3): 232-240 [Abstract] ( 571 ) [HTML 1KB] [ PDF 4174KB] ( 408 )
241 Spatial Aware Collaborative Representation Based on Augmented Spatial Spectral Features Network
LIU Shuang, ZHANG Yong
The curse of dimensionality can be caused by directly using representation learning to classify hyperspectral image due to high dimensionality, high correlation between bands and limited samples of the hyperspectral image. For the hyperspectral image, not all spectral bands are available for specific classification tasks. Therefore, spatial aware collaborative representation based on augmented spatial spectral features network is proposed in this paper. A hierarchical spatial spectral features network is built according to the low dimensional manifolds inherent in the hyperspectral image. Features of high dimensional data are extracted by training network. Spatial aware collaborative representation algorithms are utilized for classification. Experiments on two hyperspectral remote sensing datasets, Indian Pines and Pavia University, verify the effectiveness of the proposed algorithm.
2021 Vol. 34 (3): 241-252 [Abstract] ( 324 ) [HTML 1KB] [ PDF 3244KB] ( 377 )
Researches and Applications
253 Edge Anti-Aliasing Oriented Adaptive Index Map Prediction of Screen Content
SONG Chuanming, LIU Dingkun, SUN Shiqi, LIU Dan
The prediction ability of the palette coding is limited in the edge anti-aliasing regions, since the adjacent indexes do not present equivalence relation. Therefore, an adaptive predictive coding method is proposed based on edge anti-aliasing decision and multiple directional templates. Eight 4-neighbor directional prediction templates are firstly designed by introducing the Pitteway anti-aliasing decision with area weight. Subsequently, a two-dimensional linear correlation model is exploited to describe the relationship between the index to be predicted and its four reference pixels. Finally, the least square method is applied to calculate a group of weight coefficients for each directional template. The anti-aliasing algorithm executed by the display adapter is adaptively learned, and thus the predictive coding of the index map is realized. Experiments show that the proposed method achieves a higher prediction accuracy and the adaptive prediction ability of the palette coding in the anti-aliasing regions and complex edge regions is improved.
2021 Vol. 34 (3): 253-266 [Abstract] ( 350 ) [HTML 1KB] [ PDF 1132KB] ( 254 )
267 Knowledge Base Question Answering Method Incorporating Fact Text
WANG Guangxiang, HE Shizhu, LIU Kang, YU Zhengtao, GAO Shengxiang, GUO Junjun
In natural language problems, the relationship expression in the knowledge base is diversified. Therefore, matching the answers of the knowledge base question and answer through representation learning is still a challenge. To make up the shortcomings, a knowledge base question answering method incorporating fact text is proposed. Entities, entity types and relationships in the knowledge base are converted into fact text. A pre-trained language model(BERT) is employed for representation. The vector of question and answers in low dimensional semantic space is obtained using the rich semantic mode of BERT. The answer with the closest semantic similarity to the question is matched by calculation. Experiments show that the proposed method is effective and robust in answering common simple questions.
2021 Vol. 34 (3): 267-274 [Abstract] ( 372 ) [HTML 1KB] [ PDF 855KB] ( 357 )
275 Calligraphic Chinese Characters Generation Based on Generative Adversarial Networks with Structural Constraint
YU Shushi, ZHAO Jieyu, YE Xulun, TANG Chen, ZHENG Yang

A large amount of prior composition information of Chinese characters is required for the generation of calligraphic Chinese characters. Moreover, the previous data collection is demanding work, and the scalability of the research results is easily affected. To solve this problem, a method of Chinese calligraphy characters generation based on structure constraint using conditional stack generative adversarial networks is proposed. The Chinese character handwriting extracted directly from the source Chinese character image is considered as the structure constraint condition. High-quality calligraphic Chinese characters are generated by the condition stack generative adversarial network model. A semi-supervised learning method based on pseudo target samples is proposed for the dataset lacking of calligraphic Chinese characters. Furthermore, the unseen calligraphic Chinese characters during training are generated as well. Experiments show the proposed method can generate higher-quality calligraphy Chinese characters under the premise of using a few samples of a specific style of calligraphic Chinese character dataset.
2021 Vol. 34 (3): 275-285 [Abstract] ( 547 ) [HTML 1KB] [ PDF 2108KB] ( 513 )
模式识别与人工智能
 

Supervised by
China Association for Science and Technology
Sponsored by
Chinese Association of Automation
NationalResearchCenter for Intelligent Computing System
Institute of Intelligent Machines, Chinese Academy of Sciences
Published by
Science Press
 
Copyright © 2010 Editorial Office of Pattern Recognition and Artificial Intelligence
Address: No.350 Shushanhu Road, Hefei, Anhui Province, P.R. China Tel: 0551-65591176 Fax:0551-65591176 Email: bjb@iim.ac.cn
Supported by Beijing Magtech 　Email:support@magtech.com.cn