Denlogo.gif (3771 bytes)Dendronic Decisions Limited

Advanced Computing Research -- Expertise in Machine Learning

Home Applications Technology Articles Products News Company

Applications

From its early beginnings, ALN technology has been employed in numerous successful research studies conducted in leading research institutions around the world and in a broad range of applications. Now it's being used commercially.   Here are some of the ways ALNs are being used, and some ideas about how they could be used:

Data analysis

Finance

Prediction

Control and rehabilitation

3D Models and robotics

Communications

References

Analysis Applications

The use of Adaptive Logic Networks for data analysis makes use of their ability to recognize patterns by extracting useful information from noisy data. As a data analysis tool, the results are used to provide classifications or a continuous real-valued output that is functionally related to the input variables.

Estimating Fat Content in Beef1

The market value of beef is currently based on determining its grade by subjectively evaluating the amount of intramuscular fat or marbling. Highly trained and skilled human graders determine marbling levels by visually inspecting a cross-sectional area of the ribeye muscle. According to one U.S. study, grading errors by humans can be as high as 20%, with the average national error in assigning beef grades at 7.3%. A comparison made with a national panel of highly qualified beef graders had an error of 5.5%. McCauley et al. (1994) state:

Since the quality grade is one of the most important factors in determining carcass value, the beef industry needs objective methods to measure the quality of beef carcasses.

Research was conducted in the Departments of Agricultural Engineering at Purdue University and at Texas A&M University on the use of Dendronic Decisions’ Adaptive Logic Networks for predicting the intramuscular fat in beef. In order to obtain an objective measure of fat content ultrasonic images of both slaughtered and live animals were used. The results show a mean error of 0.94% on slaughtered and 0.83% on live animals. In comparison, a statistical model presented in a previous study2 for the same data predicted with a mean error of 1.38% fat of slaughtered and 1.37% on live animals. The study’s authors (McCauley et al. 1994) conclude that:

Results showed that Adaptive Logic Networks perform better than any fat prediction method for beef ultrasound images to date and are a viable alternative to statistical techniques.

Measuring odour concentration with an artificial nose10

Dr. Guoliang Qu and others working at the University of Alberta and subsequently at the Alberta Research Council have developed a system for measuring odor concentration using an artificial nose.  This system could help to provide an objective measure of how bad the odour is in the area of a pig farm, for example. As hog facilities grow in size, it becomes very important to have objective information to settle disputes with the neighbors. The alternative to using an artificial nose is a costly, time-consuming process using an olfactometer and a large odour panel.

The Use of Adaptive Logic Networks in Cardiology Research3

An on-going research project at the University of Alberta has applied Adaptive Logic Networks to the problem of detecting myocardial ischemia in an electrocardiogram (ECG). The results were taken from 1,367 study subjects aged 65 and over. Two other methods were compared with ALNs in identifying ischemic episodes from 24-hour ambulatory Holter ECG recordings. Preliminary results have shown Adaptive Logic Network technology achieves slightly higher accuracy (67% versus 56% using a back-propagation neural network).

Although the standard resting ECG is a cost-effective method of detecting myocardial ischemia events, the ECG diagnostic criteria remains sub-optimal. The anticipated successful outcome of this research to produce an Adaptive Logic Network based diagnostic could have significant impact on current medical diagnostic procedures.

Back to Top

Prediction and Forecasting Applications

Prediction, or forecasting, is the ability of the system to predict future values or outcomes based on currently known input values. The approach is for the system to learn the relationship between the input and output variables from historical data presented during a learning phase.

Predicting Compressor Unit Failure4

A feasibility study was conducted in which an Adaptive Logic Network was trained to predict failures of turbine-driven compressor units from a large database of sensor measurements. Predictions were made based on the statistical properties of the measurements and the associated failure type. Cost savings are possible by using a predictive maintenance strategy instead of a run-to-failure or scheduled-maintenance strategy. The results demonstrate the feasibility of predicting compressor failures several hours, and in some cases days, in advance of the actual shutdown time.

Magnetic Substorm Prediction5

A magnetic substorm is a disturbance of the Earth’s magnetic field partially caused by solar wind. In 1989, a magnetic substorm caused Quebec Hydro, a large power company supplying electricity to the province of Quebec and parts of the United States, to shut down. Magnetic substorms have also been known to cause problems with satellite systems resulting in large monetary losses. Research employing traditional statistical methods has failed to determine useful metrics that can be used to predict oncoming magnetic storms. Using magnetic field readings taken at ground level, several parameters were computed and used to train an Adaptive Logic Network. The preliminary results indicated that correct classifications in predicting upcoming magnetic substorms were as high as 86.6% using Adaptive Logic Networks technology.

Short Term Electric Load Forecasting6

A recently completed study on Short Term Load Forecasting (STLF) by Dendronic Decisions for Edmonton Power has indicated the potential for developing a commercial software product capable of predicting the hourly power load demand in Alberta. Electric load forecasting, for periods from one to 168 hours (7 days) in advance, is used by power companies to schedule electricity-generating units with the local power grid. The industry is currently undergoing deregulation and will require power companies to self-schedule generating units and provide competitive offers to supply power. Success by power companies in this market will depend on an ability to accurately forecast load demands to minimize the costs associated with the over-commitment of generators. Over-commitment involving several generators can cause losses of hundreds of thousands of dollars per year. Adaptive Logic Networks technology has several potential advantages over the current state-of-the-art technology both in terms of the accuracy of the forecast and the ease with which the models are obtained.

"Edmonton Power generation supplies electricity to a grid serving about 2.5 million people. Our annual energy sales are about $500M.

We know that under the proposed self commitment rules for generator operation, accurate short term load forecasting can result in significant savings and increased ability to compete with other genco’s selling to our grid.

We have worked with Dendronic Decisions on utility uses for a new type of neural network technology called Adaptive Logic Networks (ALN’s). A study performed in December ’96 demonstrated that the ALN’s learned to predict utility loads based on historical weather and load data quickly and easily. In this test, data from seven weather stations was collected. The weather in Alberta is particularly unpredictable and subject to tremendous variations over short periods, a challenge for any load predictor.

The result of the test demonstrated an error of about 0.8%, a level of accuracy that has greatly enhanced our load prediction ability.

Edmonton Power is confident that the use of Adaptive Logic Networks will almost certainly enable us to make more precise short term load forecasts that will lead to lower costs for customers through more accurate generating unit dispatch management. We are working with Dendronic on the applications of the ALN technology that will improve our operations in other areas."

Doug Topping,
Vice President, Generation
Edmonton Power

Back to Top

Control Applications

Control problems typically involve controlling one or more output quantities based on the state of several input quantities to a system. A control system is a well understood model that describes the relationship between the known input and the desired output quantities. When the relationship between the input and output is not well understood, making it difficult or impossible to create the required model, machine learning technology may be used to learn the input-output relationship.

Control of a Vehicle Active Suspension Model7

An active suspension system is used in an automobile to improve both the ride and handling qualities by controlling the force an actuator strut exerts on the vehicle’s body with respect to the road surface. Since conditions can vary from smooth to very rough terrain, an adaptive real-time controller is usually chosen for the task. Dendronic Decisions performed a successful feasibility study for the Canadian Defence Department to develop such a controller. A model was developed which predicted the dynamic state of the system 10 ms into the future, allowing the controller to adjust the strut’s force to meet the predicted road surface. The results were impressive. A stable controller was demonstrated which held a sprung mass of 200 pounds motionless to within approximately two millimeters, eliminating all except 4% of the simulated road disturbance.

Control of Walking in People with Incomplete Spinal Cord Injuries 8,9

People with spinal cord injuries (SCI) are generally at least partially paralyzed and often unable to walk. Some persons with incomplete SCI are able to manually control electrical stimulation that acts upon the nerves and muscles of a paralyzed leg and allows functional walking. With the aid of crutches or a mobile walker for support, the stimulation is activated by pressing a hand switch installed on the walking aid. However, the manual hand switch is not always appropriate for incomplete quadriplegics or stroke victims. Furthermore, the repetitive voluntary action required of a hand switch could introduce variability and cause delays.

A successful five-year research study conducted at the University of Alberta has resulted in the development of an experimental walking prosthesis that uses Adaptive Logic Networks. The system learns, from either an experienced technician or the patient, the stimulation control needed for walking. In fact, by monitoring sensor outputs, the system was also capable of predicting the person’s walking intention a full two seconds in advance, thereby providing the subject with early feedback about impending stimulation. These experimental systems may help to reestablish much of the skill needed by patients to walk reasonably long distances and enter wheelchair-inaccessible places.

3D Models

Robot Navigation11-15

3D models are required to answer the question: what part of this 3D space is occupied.  In a contract for the Defense Research Establishment Suffied, Dendronic created a robot with a single camera mounted on the end of an arm on the top.   By scanning the environment, the camera moved sufficiently to collect depth information about the space it was in so it could navigate without bumping into things.   Models that use a grid of points and assign numbers at each point to represent occupancy (0 = nothing here, 1= solid object here) can occupy a lot of storage space.   In addition, the processing of this information to answer questions about navigability is time-consuming.  A piecewise linear model using ALNs is both economical in space and computing time.

Synthetic Environments, Visualization Gaming, Movies

Dendronic has very recently used high-quality depth data provided by the Defense Research Establishment Suffield to develop a 3D model of an indoor scene.   This work is illustrated by the 3D modeling sample program that comes with the Dendronic Learning Engine.  3D models are useful for planning the construction of aircraft and other vehicles which are built in such small numbers that it has to be done right the first time.  Synthetic environments offer ways of virtually testing devices and humans in realistic situations but without the cost and difficulty of carrying out real tests.  As e-commerce develops, 3D models will be useful for creating virtual shopping malls.  Already 3D is important in computer gaming and movie entertainment.     Anyone who has experienced a large format 3D movie will realize how valuable 3D information could be in education, virtual travel, teleconferencing, shopping, and many other areas.

Communications

Call Admission Control in an ATM communication network.16

ALNs have been used to optimize ATM communication networks.  In order to preserve quality of service (QoS), a call admission control (CAC) system may not allow a call to be carried on the network if it would cause the QoS for other calls to fall to an unacceptable level. 

Back to Top

References

  1. J. McCauley, B. Thane and A. Whittaker. Fat Estimation in Beef Ultrasound Images Using Texture and Adaptive Logic Networks. Transactions of ASAE, Vol. 37, 1994, pages 997-1002.
  2. B. R. Thane. Prediction of intramuscular fat in live and slaughtered beef animals through processing ultrasonic images. Thesis, Texas A&M University College Station, Texas 1992.
  3. M. Polak, S. Zhou, P. Rautaharju, W. Armstrong and B. Chaitman. Adaptive Logic Network Compared with Backpropagation Network in Automated Detection of Ischemia from Resting ECG. Computers in Cardiology 1995, pages 217-220.
  4. W. Armstrong, C. Chu and M. Thomas. Feasibility of using Adaptive Logic Networks to Predict Compressor Unit Failure, Proceedings, Battelle Pacific Northwest Laboratories Workshop on Environmental and Energy Applications of Neural Networks, Richland WA, USA, March 30 - 31, 1995.
  5. J. P. Samson. Prediction of Magnetic Storms using Adaptive Logic Networks, unpublished Technical Report, University of Alberta, January 1997.
  6. C. Stroemich and M. Thomas. A Short Term Load Forecasting System Using Adaptive Logic Networks, Proceedings of the American Power Conference, April 1997.
  7. W. Armstrong and M. Thomas. Case Study: Control of a Vehicle Active Suspension Model using Adaptive Logic Networks, Handbook of Neural Computation, Emile Fiesler and Russell Beale, editors, Oxford University Press, 1996, pages G2.1:1 - 5.
  8. A. Kostov, B. Andrews, D. Popovic, R. Stein and W. Armstrong. Machine Learning in Control of Functional Electrical Stimulation Systems for Locomotion, IEEE Transactions on Biomedical Engineering, Vol. 42. No. 6, June 1995, pages 541-551.
  9. A. Kostov, W. Armstrong, M. Thomas and R. Stein. Adaptive Logic Networks in Rehabilitation of Persons with Incomplete Spinal Cord Injury, Handbook of Neural Computation, Emile Fiesler and Russell Beale, editors, Oxford University Press, 1996, pages G5.1:1 - 8.
  10. Guoliang Qu, J. Feddes, et al., Normalization of an Odour-Panel's Olfactory Response, Canadian Society For Engineering in Agriculture, Agri-Food 2000, Winnipeg, July 15-19, 2000.
  11. D. O. Gorodnichy and W. W. Armstrong, Neurocomputational Approach for Modeling Large Scale Environments from Range Data, Proc. ICSC Symp. on Neural Computation NC '2000, May 23 - 26, Berlin, Germany, CD-ROM.
  12. W. W. Armstrong and D. O. Gorodnichy, Breaking Hyperplanes to fit data, with applications to 3D world modeling and oil sand data analysis, Proc. ICSC Symp. on Neural Computation NC '2000, May 23 - 26, Berlin, Germany, CD-ROM.
  13. D. O. Gorodnichy and W. W. Armstrong, Vision-Based Occupancy Modeling, Proc. ISR 2000, Montreal,May 2000, pp. 293 - 298.
  14. W.W. Armstrong, B. Coghlan, D.O. Gorodnichy, Reinforcement Learning for Autonomous Robot Navigation, International Joint Conference on Neural Networks (IJCNN'99) , Washington DC, July 21-23, 1999.
  15. D. O. Gorodnichy and W. W. Armstrong, Building and Using Parametrically Represented Occupancy Models in Mobile Robot World Exploration, Robotics Today (publ. by Robotics International of the Society of Manufacturing Engineers), Vol 13, No. 2, 2nd quarter 2000.
  16. S. Ramaswamy, T. Ono-Tesfaye, W. W. Armstrong and P. Gburzynski, Equivalent Bandwidth Characterization for Real-time CAC in ATM Networks, Journal of High Speed Networks, 1998, pp 1-25

Back to Top

Home Applications Technology Articles Products News Company

Send mail to with questions or comments about this web site.
Copyright © 2003 Dendronic Decisions Limited. All rights reserved.
Date Modified: March 1, 2003.