Back to home PRODUCTS






Artificial Neural Networks mimic our brain's own problem solving process. Naturally, there are as many opinions as there are people voicing them as of how much exactly can we say about the brain... However, after about 30 years of research and development, we enjoy practical results that will convince even the most skeptical critics.

Just as we apply knowledge gained from previous experiences to solving new problems, an artificial NN "looks" at a set of answers to previously solved examples to build a system of "connections" that makes decisions, classifications, forecasts, and finds solutions to problems in a non-linear manner. The problems that best lend themselves to an artificial NN solution are those which do not have precise computational answers but which require some sort of pattern recognition. There are two modes of operation for any artificial NN - the learning phase and the recall phase. During the learning phase, we would "show" the NN multiple sets of historical problem data in which the outcomes are known, kind of a "case study", or sample problems and solutions created with the help of domain knowledge experts. During this process of "learning", the NN adjusts the strength of its internal connections, and once the training is completed, the network should be able to classify or predict based on new inputs. The more historical data is used in the learning phase, the better the resulting network at finding correlations and solving new problems. During the recall phase, the new problem’s parameters are being propagated within the artificial NN structure and a solution is being generated, based on what the network "knows" from its training. The recall phase takes almost no time, as only one set of parameters is being considered for the internal connections of the trained network.

There are numerous types of artificial NN that can be divided into categories depending on what their structure is. No matter what the type of the particular NN, they all have layers of nodes, each of which is "connected" to the nodes in the neighboring layer. These "connections" are weights which are applied to values passed from one node to the next. Within the computer, these "connections" are represented by a set of equations that are being solved sequentially for all parameters (nodes and corresponding weights). The more nodes the network consist of, the more the equations that have to be solved. Insufficient computer power is the primary reason why even nowadays we can only run very simple artificial NN. Parallel processing and advanced processing techniques decrease the required time and open new horizons to the wide applicability of artificial NN in everyday life.

On a cognitive level, the NN "learns" by adjusting its node interconnection weights. The answers the network is producing are repeatedly compared with the correct answers and on each such comparison, the connecting weights are adjusted slightly in the direction of the correct answers. In other words, the more the network is used, the better it becomes at solving the problems it was trained to solve.

Eventually, if the problem can be learned, a stable set of weights adaptively evolves and will produce good answers for all the sample decisions and predictions. At this point of time, only minute weight adjustments will be introduced in the existing distribution scheme, so the network behavior will not change noticeably from one instance to another. The real power of NN is evident when the trained network is able to produce good results for data which the network has never "seen" before. With careful design and sufficient amount of historical data, accuracy rates of 99.9% are achievable.

Check a real-life application - our HVAC systems to see how far can we go. Should your organization have any requirements in this exciting area of the technology, we will be pleased to assist you in any way in achieving your goals. For more information, please contact us at your convenience.

©1995-2004 Control Systems Unlimited. All Rights Reserved.