A Knowledge Transfer Framework Based on Deep-Reinforcement Learning for Multistage Construction Projects

Construction projects can produce excessive construction data using intelligent equipment. Deep learning algorithms can harness these data, discovering knowledge that can effectively enhance project execution and performance. However, these algorithms' efficacy hinges on the availability of data, often presenting a challenge at project initiation when data are scarce or absent. Although this issue can be alleviated through knowledge transfer algorithms, there is still a lack of a method for selecting knowledge-source domains that can adapt to data distributions. Furthermore, the long durations and dynamic environments of construction projects require timely updates to a project's knowledge base, which is often ignored by current knowledge-management practices. In this article, we introduce a multistage, online knowledge transfer framework tailored to three distinct stages of construction projects: the initiation, data-emergence, and data-rich stages. First, the framework transforms source-project data into target-project data, alleviating the initial data deficit. Subsequently, it utilizes a combination of multiple similarity metrics and a deep-reinforcement learning model to adaptively select source domains with minimal distribution disparities, improving the effect of knowledge transfer. Finally, it integrates concept-drift algorithms and constructs knowledge-source discrimination rule to automate the selection of knowledge sources and schedule updates, enabling dynamic knowledge transfer in construction projects.