APPLICATIONS OF ARTIFICIAL INTELLIGENCE IN PROJECT MANAGEMENT By Axel Gaarslev
Technical University of Denmark
Department of Construction Management
BACKPROPAGATION NEURAL NETWORKS
The Department of Construction Management at the Technical
University of Denmark has for a number of years beers involved in developing different toots baaed on artificial intelligence techniques and most of the toots have been developed in close collaboration with the construction industry in Denmark. The teds developed have been different expert systems, knowledge-based systems and neural networks, each designed to offer decision support to a specific type of problem. This paper will describe a small segment of these tools, as the paper will only focus on the latest and probably the most promising technology: neural networks. Only neural networks based on simple, standard software (Brainmaker Professional from California Scientific Software) for standard PC’s - affordable to industry and easy to use by the project manager himself - will be covered.
The paper will give a short introduction to the technology, describe in some detail an application for analyzing polluted sites and based also on further cases from practice finally offer some general conclusions on the potential to the construction industry of this kind of technology.
The human brain is made up of billions of cells called neurons. Each of these ceils is like a very small computer with extremely limited capacity - yet connected together, these cells form the most intelligent system known. Neural networks area class of computer systems formed from simulated neurons, connected to each other in a network simulating the way, that we believe the brain’s neurons are connected.
The networks in this study are based on the so called feedforward, backpropagation algorithm. In this algorithm learning is simulated in much the same way, as we think people learn, by examples and repetition - association. It is not programmed by rules etc, but it is trained . that is, when the network sees an input A, or something like it, it responds with output B, or something like it. When a neural network is being trained, it is presented input-output pairs - facts. The output portion of a fact is called a training pattern.
Each time an input is presented, the network sends back an answer of, what it thinks the output should be. If it’s wrong, the network makes corrections to itself. It means, the program goes through the list of facts,
Figure no 1: A sample neural network.
PROJECT MANAGEMENT INSTITUTE Seminar/Symposium
PITTSBURGH, PA Papers Presented: Sept. 21-23, 1992
Figure no 2: The function of a neuron.
presenting each fact once in turn and making corrections as necessary. When the entire list of facts has been presented, the program starts over at the beginning of the list. This training process is repeated until the network gets all the facts correct, if possible. Having obtained a hopefully good training pattern, the ncural network is ready to help forecast future outcomes represented by actual sets of input values. The neurons in the program are organized into layers. Figure no 1 shows an example.
The neurons are organized in three types of layers, m input layer (representing the input facts), one or more so called hidden layers, and the output layer with the facts, we are training the network to forecast. The example in figure no 1 shows a network with five input neurons, five hidden neurons in one layer, and three output neurons. The lines interrelating the neurons show, how the neurons are connected into a network, and how information flows between neurons. In this feedforward algorithm information flows only from left to right. The function in the program of a single neuron is illustrated in figure no 2.
The box represents a single neuron. It is receiving an arbitrary number of inputs, in the example 5 inputs. Neurons send a single...
Please join StudyMode to read the full document