Peng Li
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Peng Li has a total of 16 co-authors. Here are the top co-authors by number of shared publications:
- Shaofei Chen
College of Intelligence Science and Technology, National University of Defense Technology, Changsha, China
3 shared publications - Xiaolong Su
College of Intelligence Science and Technology, National University of Defense Technology, Changsha, China
3 shared publications - Ao Ma
College of Intelligence Science and Technology, National University of Defense Technology, Changsha, China
2 shared publications Haibin Xie
College of Intelligence Science and Technology, National University of Defense Technology, Changsha, China
2 shared publications- Haibin Xie
National University of Defense Technology
2 shared publications Jing Chen
2 shared publications
Latest Publications
- Open Access
- 8 May 2026
- Published by Springer Nature in Complex & Intelligent Systems
- Vol. 12 (8)
Mia-grl: mutual information aligned generative reinforcement learning with missing data
- 8 May 2026
- Published by Springer Nature in Complex & Intelligent Systems
- Vol. 12 (8)
Open AccessMia-grl: mutual information aligned generative reinforcement learning with missing data
- Open Access
- 27 March 2026
- Published by Elsevier in SSRN Electronic Journal
Distributional weighted multi-agent policy learning method with adaptive risk attitudes
In multi-agent cooperation tasks, the uncertainty of the external complex environment often leads to the randomness of rewards and state transitions, which makes the multi-agent reinforcement learning method based on risk-neutral attitude decision-making face significant risks. Traditional value distribution reinforcement learning adapts to environmental risks by modeling Q-value as distributions but ignores the dynamic change of environmental risks and the change of agents' decision-making attitude when risks change. Therefore, we propose a distributional weighted QMIX algorithm with adaptive risk attitudes (DWMIX), in which agents can adaptively adjust the decision attitude to adapt to changes in environmental risks, such as adopting risk-seeking, neutral, or averse attitude selecting actions, to find more effective exploration policy. Specifically, at the joint level, DWMIX first uses an implicit quantile network to embed environmental risk into learning the optimal joint action value distribution, and propose the expected utility Individual-Global-Max criterion for utility value decomposition. Secondly, at the individual level, DWMIX learns risk-sensitive policy with adaptive risk attitudes through the option framework and the utility value distribution, to reveal the internal change characteristics of risk attitude when agents choose policy in the cooperation scenario. Finally, we constructed several risk settings in three challenging benchmark cooperation scenarios and conducted extensive experiments, demonstrating the performance advantages of the DWMIX algorithm.
- 27 March 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessDistributional weighted multi-agent policy learning method with adaptive risk attitudes
In multi-agent cooperation tasks, the uncertainty of the external complex environment often leads to the randomness of rewards and state transitions, which makes the multi-agent reinforcement learning method based on risk-neutral attitude decision-making face significant risks. Traditional value distribution reinforcement learning adapts to environmental risks by modeling Q-value as distributions but ignores the dynamic change of environmental risks and the change of agents' decision-making attitude when risks change. Therefore, we propose a distributional weighted QMIX algorithm with adaptive risk attitudes (DWMIX), in which agents can adaptively adjust the decision attitude to adapt to changes in environmental risks, such as adopting risk-seeking, neutral, or averse attitude selecting actions, to find more effective exploration policy. Specifically, at the joint level, DWMIX first uses an implicit quantile network to embed environmental risk into learning the optimal joint action value distribution, and propose the expected utility Individual-Global-Max criterion for utility value decomposition. Secondly, at the individual level, DWMIX learns risk-sensitive policy with adaptive risk attitudes through the option framework and the utility value distribution, to reveal the internal change characteristics of risk attitude when agents choose policy in the cooperation scenario. Finally, we constructed several risk settings in three challenging benchmark cooperation scenarios and conducted extensive experiments, demonstrating the performance advantages of the DWMIX algorithm.
- Open Access
- 16 October 2025
- Published by MDPI AG in Electronics
- Vol. 14 (20) , 4064
Robust Offline Reinforcement Learning Through Causal Feature Disentanglement
Offline reinforcement learning suffers from critical vulnerability to data corruption from sensor noise or adversarial attacks. Recent research has achieved a lot by downweighting corrupted samples and fixing the corrupted data, while data corruption induces feature entanglement that undermines policy robustness. Existing methods fail to identify causal features behind performance degradation caused by corruption. To analyze causal relationships in corrupted data, we propose a method, Robust Causal Feature Disentanglement(RCFD). Our method introduces a learnable causal feature disentanglement mechanism specifically designed for reinforcement learning scenarios, integrating the CausalVAE framework to disentangle causal features governing environmental dynamics from corruption-sensitive non-causal features. Theoretically, this disentanglement confers a robustness advantage under data corruption conditions. Concurrently, causality-preserving perturbation training injects Gaussian noise solely into non-causal features to generate counterfactual samples and is enhanced by dual-path feature alignment and contrastive learning for representation invariance. A dynamic graph diagnostic module further employs graph convolutional attention networks to model spatiotemporal relationships and identify corrupted edges through structural consistency analysis, enabling precise data repair. The results exhibit highly robust performance across D4rl benchmarks under diverse data corruption conditions. This confirms that causal feature invariance helps bridge distributional gaps, promoting reliable deployment in complex real-world settings.
- 16 October 2025
- Published by MDPI AG in Electronics
- Vol. 14 (20) , 4064
Open AccessRobust Offline Reinforcement Learning Through Causal Feature Disentanglement
Offline reinforcement learning suffers from critical vulnerability to data corruption from sensor noise or adversarial attacks. Recent research has achieved a lot by downweighting corrupted samples and fixing the corrupted data, while data corruption induces feature entanglement that undermines policy robustness. Existing methods fail to identify causal features behind performance degradation caused by corruption. To analyze causal relationships in corrupted data, we propose a method, Robust Causal Feature Disentanglement(RCFD). Our method introduces a learnable causal feature disentanglement mechanism specifically designed for reinforcement learning scenarios, integrating the CausalVAE framework to disentangle causal features governing environmental dynamics from corruption-sensitive non-causal features. Theoretically, this disentanglement confers a robustness advantage under data corruption conditions. Concurrently, causality-preserving perturbation training injects Gaussian noise solely into non-causal features to generate counterfactual samples and is enhanced by dual-path feature alignment and contrastive learning for representation invariance. A dynamic graph diagnostic module further employs graph convolutional attention networks to model spatiotemporal relationships and identify corrupted edges through structural consistency analysis, enabling precise data repair. The results exhibit highly robust performance across D4rl benchmarks under diverse data corruption conditions. This confirms that causal feature invariance helps bridge distributional gaps, promoting reliable deployment in complex real-world settings.
- Open Access
- 11 November 2024
- Published by Springer Nature in Complex & Intelligent Systems
- Vol. 11 (1) , 1-15
Moor: Model-based offline policy optimization with a risk dynamics model
- 11 November 2024
- Published by Springer Nature in Complex & Intelligent Systems
- Vol. 11 (1) , 1-15
Open AccessMoor: Model-based offline policy optimization with a risk dynamics model
- 6 September 2024
- Published by Elsevier in Knowledge-Based Systems
- Vol. 304
VAOS: Enhancing the stability of cooperative multi-agent policy learning
- 6 September 2024
- Published by Elsevier in Knowledge-Based Systems
- Vol. 304
VAOS: Enhancing the stability of cooperative multi-agent policy learning
- 11 April 2024
- Published by Elsevier in Neurocomputing
- Vol. 587
An enhanced structural developmental neural network with information saturation for continual unsupervised learning
- 11 April 2024
- Published by Elsevier in Neurocomputing
- Vol. 587
An enhanced structural developmental neural network with information saturation for continual unsupervised learning
- Open Access
- 11 October 2023
- Published by Springer Nature in Autonomous Intelligent Systems
- Vol. 3 (1) , 1-16
A dynamic core evolutionary clustering algorithm based on saturated memory
Because the number of clustering cores needs to be set before implementing the K-means algorithm, this type of algorithm often fails in applications with increasing data and changing distribution characteristics. This paper proposes an evolutionary algorithm DCC, which can dynamically adjust the number of clustering cores with data change. DCC algorithm uses the Gaussian function as the activation function of each core. Each clustering core can adjust its center vector and coverage based on the response to the input data and its memory state to better fit the sample clusters in the space. The DCC algorithm model can evolve from 0. After each new sample is added, the winning dynamic core can be adjusted or split by competitive learning, so that the number of clustering cores of the algorithm always maintains a better adaptation relationship with the existing data. Furthermore, because its clustering core can split, it can subdivide the densely distributed data clusters. Finally, detailed experimental results show that the evolutionary clustering algorithm DCC based on the dynamic core method has excellent clustering performance and strong robustness.
- 11 October 2023
- Published by Springer Nature in Autonomous Intelligent Systems
- Vol. 3 (1) , 1-16
Open AccessA dynamic core evolutionary clustering algorithm based on saturated memory
Because the number of clustering cores needs to be set before implementing the K-means algorithm, this type of algorithm often fails in applications with increasing data and changing distribution characteristics. This paper proposes an evolutionary algorithm DCC, which can dynamically adjust the number of clustering cores with data change. DCC algorithm uses the Gaussian function as the activation function of each core. Each clustering core can adjust its center vector and coverage based on the response to the input data and its memory state to better fit the sample clusters in the space. The DCC algorithm model can evolve from 0. After each new sample is added, the winning dynamic core can be adjusted or split by competitive learning, so that the number of clustering cores of the algorithm always maintains a better adaptation relationship with the existing data. Furthermore, because its clustering core can split, it can subdivide the densely distributed data clusters. Finally, detailed experimental results show that the evolutionary clustering algorithm DCC based on the dynamic core method has excellent clustering performance and strong robustness.
- Open Access
- 13 September 2023
- Published by MDPI AG in Electronics
- Vol. 12 (18) , 3865
Evo-MAML: Meta-Learning with Evolving Gradient
How to rapidly adapt to new tasks and improve model generalization through few-shot learning remains a significant challenge in meta-learning. Model-Agnostic Meta-Learning (MAML) has become a powerful approach, with offers a simple framework with excellent generality. However, the requirement to compute second-order derivatives and retain a lengthy calculation graph poses considerable computational and memory burdens, limiting the practicality of MAML. To address this issue, we propose Evolving MAML (Evo-MAML), an optimization-based meta-learning method that incorporates evolving gradient within the inner loop. Evo-MAML avoids the second-order information, resulting in reduced computational complexity. Experimental results show that Evo-MAML exhibits higher generality and competitive performance when compared to existing first-order approximation approaches, making it suitable for both few-shot learning and meta-reinforcement learning settings.
- 13 September 2023
- Published by MDPI AG in Electronics
- Vol. 12 (18) , 3865
Open AccessEvo-MAML: Meta-Learning with Evolving Gradient
How to rapidly adapt to new tasks and improve model generalization through few-shot learning remains a significant challenge in meta-learning. Model-Agnostic Meta-Learning (MAML) has become a powerful approach, with offers a simple framework with excellent generality. However, the requirement to compute second-order derivatives and retain a lengthy calculation graph poses considerable computational and memory burdens, limiting the practicality of MAML. To address this issue, we propose Evolving MAML (Evo-MAML), an optimization-based meta-learning method that incorporates evolving gradient within the inner loop. Evo-MAML avoids the second-order information, resulting in reduced computational complexity. Experimental results show that Evo-MAML exhibits higher generality and competitive performance when compared to existing first-order approximation approaches, making it suitable for both few-shot learning and meta-reinforcement learning settings.
- 27 January 2022
- Published by Springer Nature in Applied Intelligence
- Vol. 52 (10) , 11555-11568
Two-stage clustering algorithm based on evolution and propagation patterns
- 27 January 2022
- Published by Springer Nature in Applied Intelligence
- Vol. 52 (10) , 11555-11568
Two-stage clustering algorithm based on evolution and propagation patterns
- 1 January 2021
- Published by Springer Nature
- p. 759-770
DRPEC: An Evolutionary Clustering Algorithm Based on Dynamic Representative Points
- 1 January 2021
- Published by Springer Nature
- p. 759-770
DRPEC: An Evolutionary Clustering Algorithm Based on Dynamic Representative Points