Why man and machine should work together in the construction industry

Insights into our research

Planning mistakes as motivation

The construction of the new Berlin Airport (BER) and many other prestigious construction projects (e.g. the construction of the Eurotunnel or the Sydney Opera House (Hall 1982)) show that planning mistakes by people often lead to delays in construction projects. Planning mistakes were first scientifically described by Kahneman und Tversky (1977). They show that when planning projects, people tend to underestimate the time needed to successfully complete the project. The consequences of these misconceptions are usually increased costs, longer project lead times and reduced quality.

However, these misjudgements are often not due to the lack of expertise of those responsible. Rather, systemic causes are at work: The highly fragmented construction industry, the increasing complexity of construction projects and their limited time frame, result in a loss of transfer of knowledge and a lack of interfaces. An analytical overview of the project, without bias from previous projects, with constantly changing parameters, represents an enormous challenge for managers. This often results in poor management and bad decisions (Bent Flyvbjerg 2020), which lead to the planning mistakes. According to Kahneman und Tversky (1977). a key lever to counteract this problem is to enable access to distributed information from a variety of projects. 

Due to the complexity of construction projects, many factors can affect decisions that must be documented as information in a project. For example, case studies in the forecast of project duration show that simple models, such as the Bromilowsche model of 1980, are not applicable in construction practice (Magnussen 2006, Flyvbjerg 2002). 

Further research shows that methods of Artificial Intelligence make better predictions because a large number of factors from a variety of projects can be evaluated mechanically (e.g. Petruseva 2012, Dissanayaka 1999, Wei 2006). In these works mainly Artificial Neural Networks (ANN) were applied. ANNs operate according to the "black box" approach. They are not "inherently" comprehensible to humans like e.g. a decision tree. Open, transparent "glass box" approaches are required for comprehensible decision-making.

Man and machine - how good are both in comparison?

In order to compare how well man and machine make decisions, different AI models were trained using a data set. The data set contains 225 projects with five descriptive characteristics in the field of building construction and infrastructure construction (i.e. road, bridge and sewer construction) in New York. Different decision trees (e.g. Random Forest, GBT, catboost) were trained to predict the duration of construction projects and then to evaluate the training model on an unknown test data set. The predictions of the AI models were then compared with the human predicted duration. The following figure shows the comparison. It shows that in most projects the prognosis of the human being was closer to the actual duration than the machine (number of cases). 

 

The comparison between man and machine in making decisions during the strategic phase of a construction project at a high level of uncertainty shows the advantages of a data-driven analysis of distributive information: Distributive information can be handled and mechanisms of action can be demonstrated transparently and objectively for decision-making. With the addition of further features and a larger project database, the results can be further improved. It can also be assumed that with a decrease in uncertainty and an increase in information within the project, better results can be produced by the machine.

Since humans are the last decision makers and the duration of a construction project is also determined by numerous soft (i.e. non-quantifiable) factors, humans remain essential in relevant decisions. A cooperation between man and machine therefore seems to be very important in the future.

Would the planning of the Berlin airport BER have been better with AI models?

Explanable AI models can support the planning and realization of construction projects, as they can objectively evaluate a large number of parameters from a variety of projects. Risks can thus be analyzed quickly and transparently. 

Since models are trained on the basis of the respective data set and so far there is no database with thousands of documented airport projects worldwide, it is not possible to train a specific airport construction project model. Therefore, the following two questions have to be asked in research:

1. To what extent can identified causal relationships be transferred to projects of a different category, location or size?

2. How do soft factors influence the training of AI models? Examples are trust, acceptance or the collaboration of the project participants. 

Planning errors can also be found in other areas

Erroneous planning is not only to be found when forecasting the duration of major construction projects. Also in other project phases the tasks are still very complex: If individual objects are planned in the construction planning, this can lead to a missing or wrong arrangement. Due to the definition of the bill of quantities, the contract design can contain passages that can be misinterpreted or missing. Or when tracking the progress of construction work, individual elements may be forgotten. 

The machine analysis and comparison of a large number of projects and sections within the project can help to reduce these planning errors. AI models can support people, reduce search times and improve evaluations. Decisions as well as the management of construction projects can ultimately be improved. 

If you have any questions or are interested in the above mentioned contents, please contact svenja.oprach@kit.edu. The results are part of a doctoral thesis at the Institute for Technology and Management in Construction.