**1. Introduction**

A project's cost should be projected with a high degree of precision; however, making a conceptual cost estimate is challenging at this time due to a lack of data [1]. Construction companies require an early budget estimate to assess whether this expenditure is acceptable and, hence, whether the project should be continued or abandoned. To estimate the contract value, the project manager usually calculates the direct cost, indirect cost, profit, and contingencies for project risks. As a result, performing an early risk assessment is critical. Stakeholders can make decisions and choices during the conceptual stage of the projects, which have major impacts on construction duration and costs, but this effect reduces as the project progresses through its life cycle [2]. Negative risks may result in schedule delays and expense overruns [3]. As a result, the project manager should concentrate as much as possible on the major risks [4]. Overall risk estimation suffers many challenges during the conceptual stage of a project due to the limited data provided. A major issue that develops in the early phase is the lack of effective and reliable overall risk estimation approaches. As a result, project decision-makers have begun to focus more on conceptual planning, where a thorough cost analysis is a critical component in achieving the project's objectives [5]. The goal of the estimation process is to make sure that the contract plans and specifications match the cost of completing the project [6].

Developing cost estimation models, both in the planning stage and in the conceptual stages of a project, has been the subject of a lot of research. Despite the importance of overall

**Citation:** Badawy, M.; Alqahtani, F.K.; Sherif, M. A Multilayer Perception for Estimating the Overall Risk of Residential Projects in the Conceptual Stage. *Buildings* **2022**, *12*, 480. https: //doi.org/10.3390/buildings12040480

Academic Editors: Srinath Perera, Albert P. C. Chan, Dilanthi Amaratunga, Makarand Hastak, Patrizia Lombardi, Sepani Senaratne, Xiaohua Jin and Anil Sawhneyand

Received: 17 March 2022 Accepted: 11 April 2022 Published: 13 April 2022

**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

**Copyright:** © 2022 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

<sup>1</sup> Structural Engineering Department, Ain Shams University, Cairo 11517, Egypt; mohamed\_badawy@eng.asu.edu.eg

project risk, there is a gap in the research on assessing the overall risk in the conceptual stage of a project when there is insufficient project information, necessitating the development of a model for predicting the overall risk at the conceptual stage of a project [7]. This research uses artificial neural networks to construct a model for evaluating the overall risk in residential projects, based on a few characteristics that may be easily recognized at a conceptual stage with acceptable accuracy. In other words, this research does not investigate the assessment of individual positive or negative risk variables, but rather the classification of the overall risk of residential projects at a conceptual stage based on the influence on project cost.

As there is a gap in developing a method to estimate the total risk in the conceptual stage, this research aims to propose a model to predict the overall risk of residential buildings at the conceptual stage using an artificial neural network with a multilayer perception.

#### **2. Literature Review**

Financial risks are regarded as the most significant risks in construction projects in Egypt and Saudi Arabia, followed by design, political, and construction risks [8]. Lack of money, a tight deadline, design revisions, insufficient information on sustainable design, and a weak definition of sustainable scope were the top hazards to sustainable building projects in the UAE [9]. At the planning stage, risk assessment models have been proposed, using an artificial neural network such as risk assessment in Saudi Arabian building projects [10] or the construction of an expressway [11]. Another model, based on system dynamics and discrete event simulation, was proposed to evaluate the impact of risk factors on project schedules in infrastructure projects [12]. Al-Tabtabai and Alex (2000) provided an ANN-based model for predicting project cost escalation owing to political concerns and the average error was 7% [13]. A model using the Bayesian Belief Network was developed to assess and enhance the implementation of residential construction projects. Improper construction procedures and poor communication were the top risk factors [14]. At a conceptual stage, the most important aspects that determine the overall risk are the use of risk management processes, the entire project duration, contract cost, and contract type [15]. The most essential criteria in the tender process are price, the scope of work, and technical resources [16].

At the conceptual stages, there is a research gap in estimating overall risk. As a result, an essential point to consider is what proportions of errors are acceptable in any model assessing the overall risk. For every equation or model, there is a rate of error, but how can this ratio be judged, meaning how one can determine if the model is accepted or not. It is not fair to judge the error rate of a model in the conceptual stage, where there is not enough information, to the error rate of a model in the design stage where there is sufficient information. It is expected that the error rate is less in the design stage than at the conceptual stage. Therefore, the error rate of any model must be compared with the extent of errors in the same stage. Since there is no research that deal with estimating the total risk of the project in the conceptual stage, the error rate in the proposed model for calculating the total risk of the project in the conceptual stage was compared with the acceptable range of the rates of estimation models for cost estimation models in the conceptual phase. Hence, to estimate the allowable range of percentage errors, the authors relied on a review of past studies on cost estimation at a conceptual stage.

To estimate the cost of school buildings in Korea, ten factors were identified. Three models were developed to calculate the cost of the school buildings, based on 217 projects. The first model was developed using neural network techniques, while the regression analysis was used in the second model and the third model was presented using the support vector machine. The results of the neural network model showed a more accurate estimate than the results of regression analysis or the supporting vector machine models [17]. Two studies were conducted in Gaza to estimate the cost of buildings at an early stage. The first research was based on seven variables and a model was proposed based on information derived from 71 construction projects using artificial neural networks [18]. While the

second research developed a model for assessing the cost of construction projects with a high degree of accuracy and without the need for a lot of information, through the use of artificial neural networks. A database of 169 projects was collected from relevant institutions in the Gaza Strip has been adopted. The artificial neural network model has eleven factors as independent inputs [19]. A study to predict the cost of construction projects at the conceptual stage in Taiwan using ten parameters. The research suggested the utilization of the evolutionary fuzzy neural inference model to enhance cost assessment accuracy. The proposed model was relied on eleven factors [20]. In Egypt, a model to assess the cost of a residential building at an early stage using the artificial neural network and data obtained from 174 residential projects. The proposed model depended on four parameters: number of floors, the area of the floor, type of external finishing, and type of internal finishing [21]. The costs of 136 executed projects were utilized to propose an artificial neural network model to predict the preliminary cost of construction projects in Yemen. The suggested model contained 17 factors [22]. In the United States, research was conducted on the difference in the computation of construction costs utilizing artificial neural networks by comparing nineteen variables in 20 projects [23]. In Taiwan, a study has presented a prototype for the rapid assessment of a proposal integrating a probabilistic cost sub-model and a multi-factor assessment sub-model. The cost-based sub-model concentrates on the cost divisions. While the multi-factor assessment sub-model captures the specific elements influencing the cost division. That research is based on 21 variables [24]. The eight previous studies mostly agreed on nine primary factors that can be used for cost estimates at the conceptual stage of a project. These nine parameters are floor area, number of floors, type of foundation, number of elevators, type of slab, type of exterior finishing, interior finishes, type of electromechanical works, and number of basements. Table 1 shows the different sources for each parameter.


**Table 1.** Sources of parameters influencing the cost estimation.

The symbol "√" means the corresponding research determined the corresponding factor as a key factor for identifying the cost estimate at the conceptual stage.

Traditional cost estimation strategies in construction projects are the most often used. They rely on time-consuming manual project cost estimation or Excel spreadsheets, rather than using computerized tools to estimate building costs. Soft computing strategies for conceptual-stage software development were compared by Bhatnagar and Ghose (2012). The feed-forward back propagation neural network model had a mean absolute percentage error (MAPE) of 13%, a cascaded feed-forward back propagation neural network model had a MAPE of 13.6 percent, a layer recurrent neural network model had a MAPE of 11.5 percent, and a fuzzy logic model had a MAPE of 3.9 percent. This means they accepted models up to MAPE with a 13.5% acceptance rate [25].

There is a lot of research that investigates the cost estimates at the conceptual stage. Each research proposed a model with a mean absolute percentage error. Table 2 shows the mean absolute percentage error for some of these prior studies. Which illustrated that the errors in the proposed models were ranged from 4–28.2%. This means that the maximum acceptable mean percentage error in the proposed model at the conceptual stage is 28.2%.


**Table 2.** The minimum absolute percentage errors for the previous studies.

Limited studies in the conceptual stage of risk estimation identified four criteria: use of risk management processes, duration of the entire project, total cost, and type of contract. As the total cost of the project can be estimated through nine criteria: floor area, number of floors, type of foundation, number of elevators, type of slab, type of external finishing, internal finishes, type of electromechanical works, and number of basements. Hence, the cost of the project can be replaced by these nine factors. The type of contract was not included due to research limitations, as the research is related to estimating the cost of housing projects based on a fixed price contract only. Hence, the initial list of criteria used to derive the overall project risk at the conceptual stage contained eleven criteria: floor area, number of floors, type of foundation, number of elevators, type of slab, type of exterior finish, interior finishes, type of electromechanical works, number of basements plus the use of risk management processes, and the entire project duration.

#### **3. Methodology**

This study's approach is divided into three sections. The first step in achieving the study's goal is to identify the essential components influencing the overall risk assessment by evaluating past research studies that focus on construction cost and risk estimation at a conceptual stage. As a result, eleven construction elements (or processes) were presented in the primary list. Five experts with at least 15 years of experience in the construction of residential projects were randomly selected. The primary questionnaire was presented to the experts using the Delphi technique to determine the parameters in the final questionnaire in three stages. In the first stage, experts were asked to add any missing criteria, if any, that could affect the overall risk and could be discovered in the conceptual stage. Data are collected from experts, revised, and re-sent back to the experts where they are asked to rank each criterion on a five-point Likert scale. After collecting the data from the second round, the averages are calculated and any factor that has a very low impact on estimating the overall risk of the project is removed from the final list. In the third round, the experts are asked to assess whether or not they agree with the final list. The third step is to develop the model. The model is simulated using artificial neural networks using Statistical Package for the Social Sciences (SPSS) software. The critical parameters that affect the estimation of

the overall risks at the conceptual stages are considered as inputs to the model while the output is the overall risk of the project. The model may contain one or two hidden layers. The number of neurons in the one-hidden-layer model can be three, four, or five. In models with two hidden layers, the number of neurons can be four in the first layer and three in the second, or five in the first layer and three in the second, or five in the first layer and four in the second. Thus, there are three different groups in terms of the number of neurons in each hidden layer. Hence, six models can be developed. The hyperbolic tangent function was used as an activation function for the hidden layers in six models, and the Sigmoid function was tested as an activation function for the hidden layers in six other models. Hence, twelve Multilayer Perceptron models have been identified and tested. To evaluate the performance of the model, the available data were also randomly divided 5-fold. The first fold contains 29 cases, while the subsequent folds contain 30 cases. Four folds were used to train the network in each model, while the fifth fold was used to evaluate the model. The final proposed model for estimating the overall risk in the conceptual stages is the model with the lowest mean absolute error rate.
