Weiye Wang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Weiye Wang has a total of 11 co-authors. Here are the top co-authors by number of shared publications:
- Jie Wu
Jiangnan University
4 shared publications - Ru-Ru Ma
Suzhou University of Science and Technology
3 shared publications - Chenchen Yu
Qingdao Haier Air Conditioning Electronics Co., Ltd
2 shared publications Dakui Luo
Fudan University
2 shared publicationsXue Wang
Jiangnan University
2 shared publications- Zhendong Gao
Fudan University Shanghai Cancer Center
2 shared publications
Latest Publications
- Open Access
- 20 October 2026
- Published by IOP Publishing in Physica Scripta
- Vol. 101 (42) , 425201
Reservoir-computing-based chaos synchronization and its application to medical image encryption
Chaos synchronization provides a physically interpretable route for generating common dynamical states at spatially separated terminals, but most existing schemes require an accurate model and explicitly injected actuation. This paper develops a data-driven synchronization framework based on reservoir computing that learns nonlinear state couplings directly from observed trajectories. A gray wolf optimizer calibrates six influential reservoir hyperparameters, and a single measured state is then used to reconstruct the remaining states online. Numerical studies on the Lorenz and Rössler systems show that the learned reservoir reaches stable synchronization from multiple initial conditions without adding a control term to the governing equations. The synchronized trajectory is treated as reproducible public session context rather than as a cryptographic secret. The image-protection architecture uses PBKDF2-HMAC-SHA-256, HKDF-separated keys, a fresh nonce, an authenticated header, and HMAC-SHA-256 verification to avoid circular plaintext-dependent masking. The evaluation includes quantitative comparisons, channel-imperfection tests, reservoir ablations, and expanded image-security experiments. In these expanded tests, 120 deterministic tiles from six source exemplars yield a mean ciphertext entropy of 7.9925 bits and a mean maximum absolute adjacent-pixel correlation of 0.0232 . One-bit perturbations give mean NPCR 99.6025 % and UACI 33.4446 % , with exact recovery in every accepted trial. These measurements are descriptive algorithmic diagnostics only; the dependent tiles do not represent 120 independent patients, clinical validation, or a cryptographic security proof.
- 20 October 2026
- Published by IOP Publishing in Physica Scripta
- Vol. 101 (42) , 425201
Open AccessReservoir-computing-based chaos synchronization and its application to medical image encryption
Chaos synchronization provides a physically interpretable route for generating common dynamical states at spatially separated terminals, but most existing schemes require an accurate model and explicitly injected actuation. This paper develops a data-driven synchronization framework based on reservoir computing that learns nonlinear state couplings directly from observed trajectories. A gray wolf optimizer calibrates six influential reservoir hyperparameters, and a single measured state is then used to reconstruct the remaining states online. Numerical studies on the Lorenz and Rössler systems show that the learned reservoir reaches stable synchronization from multiple initial conditions without adding a control term to the governing equations. The synchronized trajectory is treated as reproducible public session context rather than as a cryptographic secret. The image-protection architecture uses PBKDF2-HMAC-SHA-256, HKDF-separated keys, a fresh nonce, an authenticated header, and HMAC-SHA-256 verification to avoid circular plaintext-dependent masking. The evaluation includes quantitative comparisons, channel-imperfection tests, reservoir ablations, and expanded image-security experiments. In these expanded tests, 120 deterministic tiles from six source exemplars yield a mean ciphertext entropy of 7.9925 bits and a mean maximum absolute adjacent-pixel correlation of 0.0232 . One-bit perturbations give mean NPCR 99.6025 % and UACI 33.4446 % , with exact recovery in every accepted trial. These measurements are descriptive algorithmic diagnostics only; the dependent tiles do not represent 120 independent patients, clinical validation, or a cryptographic security proof.
- 11 September 2026
- Published by IOP Publishing in Physica Scripta
- Vol. 101 (37) , 375202
A reduced-input PPO method for synchronizing Chua’s chaotic circuits
This study addresses model-free synchronization of two Chua chaotic circuits when one of the three candidate actuation channels is unavailable. The method constructs a proximal policy optimization (PPO) controller from the synchronization errors and practically estimated error derivatives. The policy outputs only two bounded control signals, while the middle channel is fixed at zero. To sharpen learning near the synchronization manifold, the conventional Manhattan-distance reward is augmented by a fractional-power error term. A circuit-level drive–response schematic is supplied to clarify how the two learned actions enter the response circuit, and the numerical implementation is supplemented by solver-consistency and derivative-estimation protocols. Simulation results show that the selected two-input structure synchronizes all three circuit states. Relative to the conventional reward under identical network and actuation settings, the refined reward reduces residual error and settling time across the reported nominal, Gaussian-noise, parameter-mismatch, and initial-condition tests. Phase portraits, paired state time series, error surfaces, and control-effort results further demonstrate contraction of the response trajectory toward the drive attractor. These findings support reduced-input PPO as a flexible numerical synchronization strategy, while physical prototyping remains necessary before claims of hardware-level performance can be made.
- 11 September 2026
- Published by IOP Publishing in Physica Scripta
- Vol. 101 (37) , 375202
A reduced-input PPO method for synchronizing Chua’s chaotic circuits
This study addresses model-free synchronization of two Chua chaotic circuits when one of the three candidate actuation channels is unavailable. The method constructs a proximal policy optimization (PPO) controller from the synchronization errors and practically estimated error derivatives. The policy outputs only two bounded control signals, while the middle channel is fixed at zero. To sharpen learning near the synchronization manifold, the conventional Manhattan-distance reward is augmented by a fractional-power error term. A circuit-level drive–response schematic is supplied to clarify how the two learned actions enter the response circuit, and the numerical implementation is supplemented by solver-consistency and derivative-estimation protocols. Simulation results show that the selected two-input structure synchronizes all three circuit states. Relative to the conventional reward under identical network and actuation settings, the refined reward reduces residual error and settling time across the reported nominal, Gaussian-noise, parameter-mismatch, and initial-condition tests. Phase portraits, paired state time series, error surfaces, and control-effort results further demonstrate contraction of the response trajectory toward the drive attractor. These findings support reduced-input PPO as a flexible numerical synchronization strategy, while physical prototyping remains necessary before claims of hardware-level performance can be made.
- Open Access
- 12 June 2026
- Published by IOP Publishing in Physica Scripta
- Vol. 101 (23) , 235209
A dual-critic network architecture for personalized tumor treatment under a dynamical tumor model
Deep reinforcement learning (DRL) shows immense promise for devising dynamic, personalized dosing strategies for complex diseases like cancer. However, the inherent ‘double-edged’ nature of drugs remains a challenge, which possess both therapeutic efficacy and toxicity. Consequently, an optimal therapeutic strategy must balance these competing objectives rather than merely maximizing efficacy. Conventional DRL algorithms often scalarize these conflicting goals into a single reward signal, a process that blurs the learning objective and may lead to suboptimal policies. To overcome this limitation, we introduce a novel framework: the Risk-Aware Advantage Actor-Critic (RA-A2C). This framework disaggregates the feedback signal into distinct ‘reward’ and ‘risk’ components, processed in parallel by specialized reward-critic and risk-critic networks. Such a design equips each critic with a well-defined, singular objective, allowing for more accurate guidance for the actor. We also incorporate a risk preference coefficient to offer a simple and interpretable mechanism for modulating the trade-off between reward and risk. In comparative experiments against multiple state-of-the-art DRL baselines, the proposed RA-A2C model demonstrated faster convergence and superior treatment performance. RA-A2C achieved an 89.80% tumor reduction rate with the highest drug utilization rate (0.70) while requiring substantially less drug than SAC. Moreover, the proposed method maintained robust performance across mild and severe scenarios, achieving tumor reduction rates of 84.5% and 75.5%, respectively. A sensitivity analysis further showed that the risk preference coefficient can flexibly adjust treatment styles from ‘aggressive’ to ‘conservative,’ providing a flexible framework for personalized precision medicine.
- 12 June 2026
- Published by IOP Publishing in Physica Scripta
- Vol. 101 (23) , 235209
Open AccessA dual-critic network architecture for personalized tumor treatment under a dynamical tumor model
Deep reinforcement learning (DRL) shows immense promise for devising dynamic, personalized dosing strategies for complex diseases like cancer. However, the inherent ‘double-edged’ nature of drugs remains a challenge, which possess both therapeutic efficacy and toxicity. Consequently, an optimal therapeutic strategy must balance these competing objectives rather than merely maximizing efficacy. Conventional DRL algorithms often scalarize these conflicting goals into a single reward signal, a process that blurs the learning objective and may lead to suboptimal policies. To overcome this limitation, we introduce a novel framework: the Risk-Aware Advantage Actor-Critic (RA-A2C). This framework disaggregates the feedback signal into distinct ‘reward’ and ‘risk’ components, processed in parallel by specialized reward-critic and risk-critic networks. Such a design equips each critic with a well-defined, singular objective, allowing for more accurate guidance for the actor. We also incorporate a risk preference coefficient to offer a simple and interpretable mechanism for modulating the trade-off between reward and risk. In comparative experiments against multiple state-of-the-art DRL baselines, the proposed RA-A2C model demonstrated faster convergence and superior treatment performance. RA-A2C achieved an 89.80% tumor reduction rate with the highest drug utilization rate (0.70) while requiring substantially less drug than SAC. Moreover, the proposed method maintained robust performance across mild and severe scenarios, achieving tumor reduction rates of 84.5% and 75.5%, respectively. A sensitivity analysis further showed that the risk preference coefficient can flexibly adjust treatment styles from ‘aggressive’ to ‘conservative,’ providing a flexible framework for personalized precision medicine.
- 5 August 2025
- Published by IOP Publishing in Physica Scripta
- Vol. 100 (8)
Chemotherapy simulation of dynamical tumor model based on a deep reinforcement learning algorithm
Tumor treatment efficacy relies heavily on precise drug dosage control, yet traditional methods often depend on empirical formulas and lack adaptability to dynamic tumor microenvironments, risking overdose. Our investigation adopts a deep reinforcement learning (DRL) approach for real-time drug dosage control. Specifically, we employed the Advantage Actor-Critic (A2C) algorithm to perform real-time drug dosage adjustment, aiming to eliminate target cells while minimizing drug usage. Experimental results demonstrate that our approach effectively eradicates cells, and the introduction of a novel reward function further reduces the required drug dosage. Additionally, we conducted supplementary experiments showing that the model exhibits strong robustness across different tumor microenvironments. Finally, we found that models trained in noisy environments respond more effectively to noise compared to those trained in noise-free settings.
- 5 August 2025
- Published by IOP Publishing in Physica Scripta
- Vol. 100 (8)
Chemotherapy simulation of dynamical tumor model based on a deep reinforcement learning algorithm
Tumor treatment efficacy relies heavily on precise drug dosage control, yet traditional methods often depend on empirical formulas and lack adaptability to dynamic tumor microenvironments, risking overdose. Our investigation adopts a deep reinforcement learning (DRL) approach for real-time drug dosage control. Specifically, we employed the Advantage Actor-Critic (A2C) algorithm to perform real-time drug dosage adjustment, aiming to eliminate target cells while minimizing drug usage. Experimental results demonstrate that our approach effectively eradicates cells, and the introduction of a novel reward function further reduces the required drug dosage. Additionally, we conducted supplementary experiments showing that the model exhibits strong robustness across different tumor microenvironments. Finally, we found that models trained in noisy environments respond more effectively to noise compared to those trained in noise-free settings.