Can neural networks benefit public sector projects?

by Simaan AbouRizk

Three things to know about neural networks:

1. Neural networks are a form of artificial intelligence that can find hidden relationships between data to predict certain outcomes. These networks learn to make predictions based on input data. They “learn,” or are trained, from existing data sets, by establishing analytical relationships between certain features and outcomes. They are “taught” to predict output based on the input given them, e.g., from past projects.

2. Neural network analysis can be useful in many public sector contexts, including understanding the relationship between crime and street lighting, estimating property assessment values, or estimating construction costs. For example, I applied neural networks to a City of Edmonton tunnelling project, at the preliminary design phase, to develop an estimation aid. Data input included cost per meter for the tunnel and the following properties: depth, length, diameter, good/bad geotechnical conditions, and features from 20+ past projects. The neural network was trained to forecast cost per linear meter of tunnel when the specific features were present. This method has also been tested on a number of other projects.

3. When a designer wants to estimate the cost of a project, a neural network can do a quick approximation, with the click of one button, and with the input of a few tunnel features like size, depth, location, ground conditions, etc. Neural networks are a simple, quick and cost-effective method for estimation. 

Three myths about neural networks:

Myth 1: Neural networks are intelligent.

Reality: They are not intelligent. They “learn” by analyzing multiple sets of data and establishing relationships between certain factors and certain outputs. They simply provide a method of establishing analytical relationships, like statistical regression, but are easier to use and more accurate.

Myth 2: Neural networks always provide the correct answer.

Reality: A neural network can only forecast based on what it has seen (the data that it has been fed). If you input data with features outside of the boundaries of the data it was trained on, it will give you incorrect predictions.

Myth 3: It takes an enormous amount of effort to build a neural network.

Reality: Neural networks are very simple and cost-effective to build. There are many tools available that can assist in application of neural networks.

Simaan AbouRizk holds an NSERC Senior Industrial Research Chair in Construction Engineering and Management and a Canada Research Chair in Operation Simulation at the Department of Civil and Environmental Engineering, University of Alberta, where he is a Professor in the Hole School of Construction Engineering. He received the ASCE Peurifoy Construction Research Award in 2008. He was elected fellow of the Royal Society of Canada in 2013.