Zhongsheng Hua

This paper investigates the dynamic coopetition and capacity-sharing strategies between an Integrated Manufacturer and a Developer under R&D uncertainty, focusing on the governance of strategically scarce capacity. By constructing a two-stage game model, we analyze how government intervention and risk-hedging mechanisms influence the allocation of idle strategically scarce capacity in innovation-driven industries. The findings reveal a two-sided paradoxical behavioral pattern: in the low-probability R&D interval, rather than relying on safe contract manufacturing, the Integrated Manufacturer counter-intuitively reduces collaborative duration to aggressively gamble on its immature product. Conversely, in the high-probability R&D interval, where conventional wisdom predicts an aggressive pivot to self-production, the manufacturer paradoxically extends or maintains the contract manufacturing duration, driven by the partner’s full cost-sharing incentive mechanism. Furthermore, To maximize the total supply output of strategically scarce resources during collaboration, we uncover a non-linear ‘counterproductive subsidy trap’ and propose a binary ‘critical mass’ policy rule: governments should either withhold subsidies entirely or commit sufficient funding to bypass the supply deficit zone. This framework provides a theoretical foundation for managing scarcity in capital-intensive sectors such as biopharmaceuticals and semiconductors.

Data collected from different channels that describe the condition of the same object may differ in their reliability and sampling costs. For example, online sensor tracking data of an object may be less reliable but much cheaper than offline field inspection data. This provides chances of fusing multichannel data to accurately monitor the condition of an object at low costs. In the article, we formulate a model of dynamically fusing easily achieved online data with costly offline sampled data (abbreviated as DFO²). In DFO², we consider a dynamic decision process where a decision-maker observes online data and then decides whether to acquire a new piece of offline data. Offline data are costly to acquire, but it is accurate and can yield a reward in correcting errors in online data. A nonlinear contextual bandit method is then proposed to estimate the expected reward of offline sampling decisions, and an offline sampling policy is obtained by maximizing the expected reward. Theoretical analysis indicates that DFO² achieves a sublinear regret bound, which means that the reward of DFO² asymptotically approaches that of the optimal policy over time. To demonstrate the wide applicability of DFO², experiments are performed across two distinct domains—healthcare and power systems. Results show that DFO² has a better performance in trading off sampling cost and information accuracy compared to the benchmarks. Comparative experiments under different conditions also reveal the robustness of the method performance. Overall, this article provides a practical framework for unifying multichannel data to realize cost-effective monitoring.

Parkinson's Disease (PD) treatment is challenging due to symptom heterogeneity and the lack of a definitive cure. Lifelong medication requires personalized treatment plans developed by physicians, but such approaches are constrained by high costs and limited physician capacity. Although deep learning (DL) methods have been explored, they lack interpretability and are restricted to numerical data inputs. In this study, we propose a novel framework that leverages large language models (LLMs) to design personalized PD treatment strategies, integrating both patient information in natural language form and external textual knowledge sources (e.g., medical guidelines). To enhance effectiveness, we use Monte Carlo Tree Search (MCTS) to refine strategies and establish a robust medication recommendation dataset. To enhance reliability and interpretability, we incorporate Retrieval-Augmented Generation (RAG) and Chain-of-Thought (CoT) reasoning within the LLM system, ensuring that each proposed strategy is accompanied by step-by-step explanations and references to similar historical cases. Experimental evaluations using the Parkinson's Progression Marking Initiative (PPMI) dataset show that our method surpasses physician-prescribed treatments, achieving an average reduction of over 1.4 points in the revised unified Parkinson's disease rating scale part III (MDS-UPDRS-III) scores. Our method also outperforms the RL-method by 1.01 points on average. Furthermore, over 43% of patients achieve more than 2 point-reduction of MDS-UPDRS-III scores. A detailed case study highlights the flexibility of LLMs in dynamically adjusting medication plans for patients at different disease stages, highlighting its potential to advance personalized PD management in real-world settings.

This research investigates the strategic decision related to original content provision and business model selection within the rapidly evolving online video industry. We develop a two-sided market model involving a video platform, users, and advertisers to analyze how the platform balances subscription and advertising revenue while offering original content. This study finds that the amount of original content provided directly influences market equilibrium, affecting the platform’s choice between a purely subscription-based model or a mixed model. The cost of original content production plays a critical role in this decision-making process. High production costs may lead the platform to favor a mixed model, offering less original content to generate both subscription and ad revenue. Lower costs, however, encourage a subscription model with more original content to boost subscribing revenue. Additionally, factors such as network externalities between users and advertisers also impact the business model choice. Importantly, the model shows that allowing free users limited access to original content can expand both subscriber and advertiser engagement, enabling platforms to enhance dual revenue streams without having to sacrifice one for the other.

This research investigates the strategic decision related to original content provision and business model selection within the rapidly evolving online video industry. We develop a two-sided market model involving a video platform, users, and advertisers to analyze how the platform balances subscription and advertising revenue while offering original content. The study finds that the amount of original content provided directly influences market equilibrium, affecting the platform’s choice between a purely subscription-based model or a mixed model. The cost of original content production plays a critical role in this decision-making process. High production costs may lead the platform to favor a mixed model, offering less original content to generate both subscription and ad revenue. Lower costs, however, encourage a subscription model with more original content to boost subscribing revenue. Additionally, factors such as network externalities between users and advertisers also impact the business model choice. Notably, the model shows that allowing free users access to a portion of original content may help platforms expand both subscription and advertising revenue without needing to choose between them.

The risk of severe complications poses a major threat to patients suffering from chronic diseases, and risk prediction models can assist with identifying patients’ risks of developing complications. Different from those static prediction models (logistic regression, decision trees, etc.) which use the patients’ cross-section data to predict whether they will suffer a complication, this research aims to provide a dynamic prediction method that comprehensively utilizes the patients’ longitudinal Electronic Health Records (EHR) to predict their complication progressions. The proposed generalized mixed prediction chain model (GMPC) takes the patient’s complication status as the response variable and takes the patient’s EHR data as the covariate. A mixed effect model is then employed to demonstrate the relationship between the response variable and the covariate. Additionally, to reduce the time delay between the historical EHR and the future complication status, GMPC constructs a prediction chain that uses the predicted value of the response variable at the previous time to support the prediction of the response variable at the next time. Internal cross-validation and external test verify the effectiveness of GMPC, and the results show that GMPC outperforms existing static prediction models and dynamic prediction models.

Coronaviruses have caused multiple global pandemics. As an emerging epidemic, the coronavirus disease relies on nonpharmacological interventions to control its spread. However, the specific effects of these interventions are unknown. To evaluate their effects, we extend the susceptible–latent–infectious–recovered model to include suspected cases, confirmed cases, and their contacts and to embed isolation, close contact tracing, and quarantine into transmission dynamics. The model simplifies the population into two parts: the undiscovered part (where the virus spreads freely—the extent of freedom is determined by the strength of social distancing policy) and the discovered part (where the cases are incompletely isolated or quarantined). Through the isolation of the index case (suspected or confirmed case) and the subsequent tracing and quarantine of its close contacts, the infections flow from the undiscovered part to the discovered part. In our case study, multisource data of the novel coronavirus SARS-CoV-2 (COVID-19) in Wuhan were collected to validate the model, the parameters were calibrated based on the prediction of the actual number of infections, and then the time-varying effective reproduction number was obtained to measure the transmissibility of COVID-19 in Wuhan, revealing the timeliness and lag effect of the nonpharmacological interventions adopted there. Finally, we simulated the situation in the absence of a strict social distancing policy. Results show that the current efforts of isolation, close contact tracing, and quarantine can take the epidemic curve to the turning point, but the epidemic could be far from over; there were still 4,035 infected people, and 1,584 latent people in the undiscovered part on March 11, 2020, when the epidemic was actually over with a strict social distancing policy.

In fault prognosis, the individual heterogeneity among degradation processes of equipment is a critical problem that decreases the reliability and stability of prognostic models. The presence of the diversity of degradation mechanisms, along with the complex temporal nature of multivariate measurements of equipment, make the existing approaches difficult to forecast the trend of health status and predict the Remaining Useful Life (RUL) of equipment. To resolve this problem, this article proposes a dual-network approach for online RUL prediction. The proposed approach predicts the RUL by constructing a recurrent neural network (RNN) and a Feedforward Neural Network (FNN) from the degradation measurements and failure occurrence data of equipment. The RNN is used to predict the evolution of degradation measurements, whereas the FNN is used to determine the failure occurrence based on the predicted measurements. Considering the individual heterogeneity problem, a novel meta-learning procedure is proposed for network training. The main idea of the meta-learning approach is to train two network generators to capture the average behavior and variation of equipment degradation, and generate dual networks dynamically tailored to different equipment in the online RUL prediction process. Numerical studies on a simulation dataset and a real-world dataset are performed for performance evaluation.

The rapid development of sensing technologies has enabled sensors embedded into systems and generated continuous online data for anomaly monitoring and protection of the system. However, due to transmission errors and environmental noises, the online measurements collected by sensors are usually inaccurate and cannot be applied directly. A small amount of precise in situ measurements, which are accurate but costly to acquire, are usually fused with online measurements for accuracy improvement. Over the past years, there are many studies concerning such data fusion approaches for obtaining reliable fused data. However, most existing works assume the availability of in situ measurements, without providing a strategy for sampling the expensive in situ measurements. As a result, they may undersample or oversample the in situ measurements that result in large fusion errors or unnecessary costs. To address this problem, this article proposes a dynamic policy for sampling the in situ measurements adaptively using the information that the online measurements provide. Specifically, the proposed policy models the dependence between in situ sampling decisions and influencing factors using a policy network, predicts the probability of recommending in situ samplings with the policy network, and guides the sampling of in situ measurements dynamically. Theoretical and experimental results both verify the effectiveness of the proposed approach. Note to Practitioners —In modern industrial systems, a large number of online measurements of equipment parameters usually can be acquired through sensors at a small cost. However, such measurements are normally corrupted with noise and not appropriate for high-accuracy monitoring of equipment. In practice, a small number of in situ measurements, which are accurate but costly to acquire, are usually collected and fused with online measurements for accuracy improvement. Many studies have been conducted on such data fusion approaches, but most of them focus on the fusion models and overlook the role of the sampling policy of in situ measurement on the fusion results. This article addresses the problem of sampling a few expensive in situ measurements to be fused with online measurements for obtaining high-accuracy and low-cost fused data. To support this objective, a dynamic sampling policy is proposed to guide the sampling of in situ measurements and balance the benefit and cost brought by in situ samplings. Experimental results on the real-world data set show the benefits of the proposed policy compared to relevant methods.