ANN techniques mentioned in the reviewed literature
| Reference | Technique | Brief description | Phase of application | Input |
|---|---|---|---|---|
| Dimitriou et al. (2018) | Feed-forward neural networks (FNN) | FNN is one of the simplest forms of ANN. Communication between neurons is only processed in one direction in this neural network. The information can be passed to the subsequent layer of neurons, but never backwards | Planning | 68 Bill-of-quantities from road bridge projects |
| Juszczyk et al. (2018), Pessoa et al. (2021) | Multilayer perceptron (MLP) | MLP is a specific case of FNN, in which every layer is a fully connected layer (FCL). A perceptron is a computational unit used for learning binary classifiers. Perceptrons have weighted input signals and produce output signals based on an activation function. The association of these units in layers creates an MLP network | Design, Planning | 129 construction projects defined by ten different variables; 1094 construction projects defined by up to four variables. |
| Moon et al. (2021b) | Recurrent neural networks (RNN) | Derived from FFNs, RNNs contain loops that allow for information to be stored and accessed by the network in the future. This information works as valuable experience in forthcoming decision-making, enabling a better performance when facing sequence-based problems (e.g., action classification over time) | Design | 4,659 sentences labelled according to five categories of information |
| Xue et al. (2020) | Convolutional neural networks (CNN) | A convolutional theoretically is an operation on two functions that produces a third. The computational model emulates this operation by stacking convolutional layers, each capable of recognising more sophisticated and complex features within the same data (usually images) | Execution | 415 expressways projects defined by five factors |
| Cao and Ashuri (2020), Jeon et al. (2021b), Cheng et al. (2020) | Long short-term memory (LSTM) | LSTM networks are improvements over an RNN. LSTM includes units that can maintain information in memory for long periods. It is possible to control when information enters the memory, when it is outputted and when it is forgotten. This is possible by using three gates: “Input”, “output” and “forget” gates. The input gate decides how much information from the last sample is stored in memory, the output gate determines how much data is transmitted to the subsequent layer and forget gates control the rate at which memory is eliminated. This structure allows for longer-term dependence analysis | Execution | 13 projects, defined by five different variables; 11,060 sentences manually labelled to rule-based classification; Cost indexes collect over 20 years |
| Juszczyk et al. (2018) | Radial basis function (RBF) | RBF networks consist of an input vector followed by a layer of RBF neurons and an output layer comprised of a set of neurons. This algorithm classifies the similarity between the input points and points from the training set (prototypes) that each neuron stores in memory. Each neuron computes the Euclidean distance between the input and its prototype. From this comparison, a similarity measure of 0 to 1 is produced. When the input is equal to the prototype, the value is 1; when they are not similar, the value drops exponentially to 0 | Design | 115 construction projects defined with ten variables |
| Reference | Technique | Brief description | Phase of application | Input |
|---|---|---|---|---|
| Feed-forward neural networks (FNN) | FNN is one of the simplest forms of ANN. Communication between neurons is only processed in one direction in this neural network. The information can be passed to the subsequent layer of neurons, but never backwards | Planning | 68 Bill-of-quantities from road bridge projects | |
| Multilayer perceptron (MLP) | MLP is a specific case of FNN, in which every layer is a fully connected layer (FCL). A perceptron is a computational unit used for learning binary classifiers. Perceptrons have weighted input signals and produce output signals based on an activation function. The association of these units in layers creates an MLP network | Design, Planning | 129 construction projects defined by ten different variables; 1094 construction projects defined by up to four variables. | |
| Recurrent neural networks (RNN) | Derived from FFNs, RNNs contain loops that allow for information to be stored and accessed by the network in the future. This information works as valuable experience in forthcoming decision-making, enabling a better performance when facing sequence-based problems (e.g., action classification over time) | Design | 4,659 sentences labelled according to five categories of information | |
| Convolutional neural networks (CNN) | A convolutional theoretically is an operation on two functions that produces a third. The computational model emulates this operation by stacking convolutional layers, each capable of recognising more sophisticated and complex features within the same data (usually images) | Execution | 415 expressways projects defined by five factors | |
| Long short-term memory (LSTM) | LSTM networks are improvements over an RNN. LSTM includes units that can maintain information in memory for long periods. It is possible to control when information enters the memory, when it is outputted and when it is forgotten. This is possible by using three gates: “Input”, “output” and “forget” gates. The input gate decides how much information from the last sample is stored in memory, the output gate determines how much data is transmitted to the subsequent layer and forget gates control the rate at which memory is eliminated. This structure allows for longer-term dependence analysis | Execution | 13 projects, defined by five different variables; 11,060 sentences manually labelled to rule-based classification; Cost indexes collect over 20 years | |
| Radial basis function (RBF) | RBF networks consist of an input vector followed by a layer of RBF neurons and an output layer comprised of a set of neurons. This algorithm classifies the similarity between the input points and points from the training set (prototypes) that each neuron stores in memory. Each neuron computes the Euclidean distance between the input and its prototype. From this comparison, a similarity measure of 0 to 1 is produced. When the input is equal to the prototype, the value is 1; when they are not similar, the value drops exponentially to 0 | Design | 115 construction projects defined with ten variables |