A human touch? How machine learning can improve project performance (fo)
Paper i proceeding, 2019
Construction projects are influenced by interrelated issues that may result in cost and/or time overruns, thus affecting the overall project performance. Therefore, a need to develop predictive models is widely highlighted, to aid in decision-making and offer guidance for corrective actions – especially when preparing for the production phase. Predictive models can utilize certain key performance indicators (KPIs). This study aims at investigating possible applications of machine learning (ML) for the development of such predictive models in construction projects, and the way these can impact project performance. Initially, a literature review about ML in the construction context is conducted. Following, two cases of developed ML predictive systems for construction project performance appraisal are presented. The first case is drawing on a productivity survey of 580 construction projects in Sweden, in which the most influential project performance factors are analyzed. The data encompasses project attributes, external influencing factors, and project organization. Statistical correlation is used to find the features that are strongly correlated with four KPIs: cost and time variance, and client and contractor satisfaction. Then, a regression analysis is performed to develop the prediction model. Technical complexity, like the level of prefabrication, are among the features affecting project performance. Moreover, human-related factors (e.g. client role, architect performance, and collaboration level), end up being highly impactful; it derives that they are the most suitable factors for predicting project success. The second case appraises a project’s constructability combined with risk analysis, via a ML model utilizing a restricted dataset of 30 diverse civil engineering projects from several different countries and with very different character; a town square, a biogas plant, road bridges and sub projects from an airport. The development built on a literature study, expert interviews, and unsupervised and supervised ML. The ML-enabled strengths of this model lie more in the novel derivation of construction project risk sources from the related body of literature, as well as the computational and not just conceptual integration of constructability and risk analysis, rather than the system’s coverage of the full corresponding context. It can be concluded that the human touch is still needed in preparing future construction projects – and even more so after the introduction of ML solutions. While ML includes human aspects, such as satisfaction and risk perception translated into concepts and variables, there is also a need for strengthening the human touch of qualified thinking for the related decision-making in construction project processes.
Supervised and unsupervised learning