Machine learning supported multiple criteria decision making for project delivery system selection

The construction industry is a backbone of the national economy throughout the world as it accounts for a sizeable proportion of most countries' gross domestic product. It is generally accepted that construction is a project-based industry, and the selection of project delivery system (PDS) is one of the most important factors affecting the success of a construction project. PDS specifies project employer's management functions during the project, and it reflects roles, responsibilities, and risks for project stakeholders. Despite considerable efforts by scholars to this end, existing research on PDS selection suffers apparent limitations, such as a lack of project characteristic analysis and a heavy reliance on expert experience. Therefore, building a comprehensive framework to select the PDS for construction projects becomes the focus of this study. In order to gain an in-depth understanding of PDS and the current state of research on PDS selection, this study analyzes the above contents through a literature review, drawing a judgment that the PDS selection problem can be split into two sub-problems (modules): Project Scheduling Problem (PSP) and Multiple Criteria Decision Making Problem (MCDM). The former attempts to explore the characteristics of the project itself, i.e., the overlap strategy between project activities, and then select the appropriate PDS, while the latter tries to summarize and utilize the historical project information, i.e., the employer's preference, project performance, etc., and to determine the suitable PDS. Therefore, this study conducts a literature review of these two sub-problems and their generic approaches. For the purpose of providing inspiration for the development of new methods to solve the above sub-problems, this study further provides a systematic literature review of MCDM problems in the construction industry and corresponding solutions. Regarding PSP, this study proposes a multiple overlapping Mode-Resource-Constrained Project Scheduling Problem-Generalized Precedence Relations (MOM-RCPSP-GPR) model to reflect the tradeoff between the overlap and the rework of activities. According to the model's characteristics, this study first develops an improved Genetic Algorithm (GA) hybrid method based on previous studies. Besides, this study redefines MOM-RCPSP-GPR to make it applicable to Priority Dispatching Rule (PDR)-based methods. Subsequently, this study presents a Deep Reinforcement Learning (DRL) method to address MOM-RCPSP-GPR. Compared to GA, DRL is more suitable for the case where individual parameters may change in a dynamic environment. Meanwhile, the shorter computation time of DRL makes it more capable of solving projects with a larger number of activities. For each of these methods, this study designs computational experiments to verify its validity; the results reflect that these methods are valid and better suited to solving MOM-RCPSP-GPR. Given that traditional PDS selection methods mainly rely on the subjective preferences of decision makers and cannot remove the influence of environmental factors, this study further proposes an advanced Case-Based Reasoning (CBR) method that integrates the nonparametric production frontier method to address these shortcomings. After a questionnaire survey, a case base is built, and a self-assessment is conducted to prove the validity of the advanced CBR method. Based on the above analysis, this study innovatively adopts a combination of advanced CBR and PSP, building a comprehensive case-based framework to handle qualitative and quantitative information. A case study based on a real-life project, an urban rail transit project funded by Multiple Development Banks (MDB), is carried out to demonstrate and evaluate the proposed framework. The results show that, compared to the industry PDS selection tools, the comprehensive framework presented in this study is complete and advanced. In summary, this study has theoretical and practical implications for the construction industry.