Home Computer Science

# PROBLEMS IDENTIFIED

In the previous section, the authors used several types of prediction networks such as BP algorithm, adaptive conjugate gr adient network, and radial basis function neural networks (RBFNNs). However, it was observed that each technique has several drawbacks, as discussed in this section.

12.3.1 DEFICIENCIES IN THE BP ALGORITHM

The previous literature shows that BP algorithm has been widely used owing to its simplicity and ease in the application [8,79]. However, this algorithm has a slow rate of learning with the number of iterations exceeding several thousand [26]. Moreover, network convergence is significantly affected by the selection of input parameters and then values along with the momentum ratios. Furthermore, the variety of input parameters has no specific rules and depends upon the type of problem [1]; (Yell, 1998). Adeli and Hung [1] were one of the first researchers who attempted to remove the deficiencies and developed a new algorithm for training of feedforward networks with a large number of layers. Powell [78] introduced further modifications in the gradient algorithm. The main problem encountered with the BP algorithm about the selection of input variables and momentum ratios has been solved by varying these parameters in a specific range leading to better convergence. The proposed algorithm was tested for image processing, and superior performance was obseived. However, this approach had the limitation that it could be applied for specific problems only. Nevertheless, it is more complicated and consumes more significant time and effort to implement it.

In the RBFNN, the mapping of the input-output parameter is performed by using a transfer function surrounding these parameters (Moody and Darken, 1989; Poggio and Girosi, 1990). In the RBFNN, the biological neurons have receptive field such that output is large when input functions are closer to the center. The topology of RBFNN includes input, hidden, and output layers arranged radially. The RRBFFS is more computationally efficient than the BP algorithm but is computationally complex, and its accuracy varies depending upon the number of layers and input parameters.

# SOLUTION TO EXISTING PROBLEMS

The problems with a different type of networks have been identified in the previous section. These problems can be solved by the application of prescribed solutions as follows

12.4.1 INTEGRATING NEURAL NETWORKS WITH OTHER ALGORITHMS

The above algorithms can be effectively integrated by combining their advantages and reducing their disadvantages. The backpropagation network and the genetic algorithm were combined by Adeli and Hung (1991b) to propose a new algorithm. The first stage begins with the acceleration of the learning rate by the adoption of feedforward networks. The corresponding weights of neural networks are decided to reduce the mean square error. The second stage starts with the adoption of the BP algorithm, and the iteration is continued until the end condition is satisfied. The concept of the hybrid algorithm has been used by several authors such as [1,49,68,72], and Topping et al. (1998).

12.4.2 FUZZY LOGIC

The fuzzy logic modeling can be subdivided into three stages. The dataset is subjected to unsupervised learning in the first stage. The second stage precedes with supervised learning and ends with the classification of the training data. The ANN and the genetic algorithm are used in combination for this purpose. After that, the activity of defuzzification and computation of the difference between the initial and final values to compute error is performed.

12.4.3 WAVELETS

The neural networks exhibit poor accuracy in case of extensive data and complicated patterns in case of traffic data collected near a central station. As estimation of construction cost involves a large number of parameters. The neural networks have observed less accuracy in case of noncoincident patterns because of the significant variation in dimensions of the training data. As a solution, one can use a noise reduction technique based on wavelet theory to inhibit unwanted fluctuation in the training data. The wavelets can be effectively used for preprocessing of the data to be fed into the neural networks, which improves the efficiency of these networks. This technique has been effectively used by several researchers such as Adeli and Samant [7], Adeli and Hung [1], and Adeli and Karim [5] for solving traffic-related problems. Several authors, Yu et al. [104] and Zhao et al. [106], determined damage location and identified it using neural networks and wavelet theory.

# PROBLEMS WITH EXISTING SOLUTIONS

The problems with the existing solutions have been presented in the following subsection.

12.5.1 PROS AND CONS OF EXISTING SOLUTIONS

The application of frizzy logic involves immense technical knowledge of input parameters for the development of fuzzy sets. The second difficulty pertains to the modeling of a fuzzy set by the description of the relationship between input parameters. This theory becomes complicated in case of problems with a large number of input parameters. Moreover, it involves a higher degree of mathematical solutions, which makes it even complex to implement.

The wavelets have a unique advantage that tune and frequency of the system can be located more accurately and can be easily adapted to solve the complex engineering problems. The wavelets have distinct advantages over Fourier transforms that they are more simple and easy to implement. Although the implementation of wavelets is simple, its design varies from problem to problem and involves lots of complexities.

However, once fuzzy sets are formulated for a specific design problem. Then, it can be implemented for similar problems with little effort.

# CONCLUSIONS

Artificial intelligence can be defined as a system specially designed to exhibit human-like behavior in the solution of complex issues. The decisions made by artificial intelligence are quick, efficient, and straightforward. Previous literature survey shows that the ANN has been extensively used for problems related to construction scheduling and forecasting. The review of prior literature shows adoption of the BP algorithm in construction management owing to its ease of application and simplicity. However, the BP algorithm had a distinct disadvantage that it was slow and required a large number of iterations to fulfill the convergence criteria. Moreover, the selection of input parameters significantly depended on the type of problem and had no specific rules. Therefore, the formulation of the generalized algorithm for a particular kind of issues becomes difficult. Nevertheless, effects with momentum ratios were also observed to satisfy the convergence criteria. Several solutions to overcome the disadvantages of the BP algorithm were prescribed: (a) integrating the BP algorithms with a genetic algorithm and (b) usage of wavelets and fuzzy logic. The proposed solutions had several disadvantages, which can be overcome using simple measures. Moreover, these techniques can be applied to a variety of problems in construction management. Nevertheless, these measures are quite complex and challenging to implement. However, these techniques can be easily and effectively implemented with the development of new software.

# FUTURE SCOPE

The present study can be extended to more aspects of construction management such as construction scheduling, construction delay, etc. The present study can be conducted with more number of variables, and different activation functions could be adopted.

# KEYWORDS

• neural networks
• construction management
• artificial intelligence
• civil engineering

 Related topics