NeuroShell Predictor 2
類神經網路預測應用軟體
Professional system to solve forecasting and estimation problems by learning historical data.
軟體代號:7141
瀏覽次數:5915
教育版
商業版
再啟動服務
遠端展示
在地教學
遠端安裝啟動服務
原廠技術服務
產品說明會
教育訓練
教學範例檔
永久授權
目前庫存
中文型錄
安裝序號
合法保證
電子英文手冊
ESD網路下載
影音教學檔
Features
The NeuroShell Predictor enables you to build powerful predictive models quickly.
NeuroShell® Predictor - This product is used for forecasting and estimating numeric amounts such as sales, prices, workload, level, cost, scores, speed, capacity, etc. NeuroShell Predictor is a simple step by step process that uses recognized forecasting methods to look for future trends in your existing data. It contains the most sophisticated prediction algorithms available today yet is designed to be extremely effective with minimum intervention by the user. The program reads and writes text files for easy access to and from other programs.
The prediction algorithms (one is a new neural network and the other is a statistical estimator driven by a genetic algorithm) are the crowning achievement of several years of research. Gone are the days of dozens of parameters that must be artistically set to create a good model without overfitting. Gone are the days of hiring a neural net expert or a statistician to build your predictive models.
Two of the most commonly heard complaints about previous prediction systems, aside from being too hard to use, are that they are too slow or that they do not accurately tell you how important each of the variables is to the model. We've taken care of those problems. That's why we have two training models from which to choose.
The first model called TurboProp2 dynamically grows hidden neurons and trains very fast. TurboProp2 models are built (trained) in 10 to 30 seconds on a 200 MHz Pentium (a matter of seconds on newer computers), compared to hours for older neural networks types. TurboProp2 reveals the relative importance of each of your inputs.
The genetic training method trains everything in an out-of-sample mode; it is essentially doing a 'one-hold-out' technique, also called 'jackknife' or 'cross validation'. If you train using this method, you are essentially looking at the training set out-of-sample. This method is therefore extremely effective when you do not have many patterns on which to train. The genetic training method takes a little longer to train. It also reveals the relative importance of each of your inputs. You will know which data you don't have to collect anymore!
The NeuroShell Predictor facilitates integration with other programs, because it uses standard text files. These files are easily imported/exported from spreadsheet programs such as Excel and Lotus®, for example.
The NeuroShell Predictor learns to make predictions by learning patterns in your data file. The Predictor allows you to select columns from your data file which will be used as inputs to the model and to select the output column which contains the values you are trying to predict.
click on image for a better picture
The NeuroShell Predictor is so easy to use that it doesn't need a manual! Instead, there is an 'Instructor' that guides you through making the predictive models. At every stage of the Instructor, our extensive help file will give you all the information you need. When you have learned from the Instructor, you can turn it off and work from the toolbar or menus. (The program does include an on-line manual that you may print yourself or just browse from your computer.)
Finally, for those who want to embed the resulting neural models into your own programs, or to distribute the results, there is an optional Run-Time Server available. Predictor models may be distributed without incurring royalties or other fees.
The NeuroShell Predictor shows you the estimated relative importance of each variable in the model.
--------------------------------------------------------------------------------
Specifications
Software Requirements
The NeuroShell Predictor is a 32-bit program that requires Microsoft® Windows® 98, Windows 2000®, XP or Windows NT® (SP3 or higher). It will not run with Windows 3.1.
Hardware Requirements
IBM® PC or compatible computer with a 486 or higher processor and 16 megabytes of RAM.
Limits
150 input variables and one output variable.
16,000 rows of data (example patterns).
NOTE: These limits are not inhibitive as they may seem for owners of large databases. Call for a more detailed explanation.
Files
ASCII text files separated by commas, spaces, tabs, or semicolons.
If your data is in a spreadsheet, simply save it as a .CSV file.
Speed
Neural nets train very fast, usually in under a minute.
The Genetic method will train very slowly on large files. This method
may be more suitable for less than 3,000 rows of data.
Statistics and Graphics
Actual vs. Predicted.
Learning level (R squared, average error, correlation, mean squared
error, root mean squared error, % in range).
Importance of Inputs.
Scatter plot.
Methodology
There are two neural network paradigms. One is a proprietary algorithm called TurboProp(TM) 2, which is NOT based on the old backpropagation algorithm. Another paradigm in the software uses an advanced variant of General Regression Neural Nets (GRNN), not to be confused with regression analysis.
--------------------------------------------------------------------------------
Features
Ability to Select Level of Generalization in Neural Training Strategy
Our neural method can actually be changed after it is trained so that it provides more or less generalization. Pressing the Advanced Button will allow you to select the level of generalization from 0% (No Enhanced Generalization) to 100% (Over Generalization). A setting of 50% is equivalent to Enhanced Generalization. The default value, when the Enhanced Generalization button is checked, is 50%.
Maximum Number of Hidden Neurons for Neural Training Strategy
You may set the number of hidden neurons to a maximum of 150 when using the Neural Training Strategy. This gives you some control over how the neural net fits data. You may even specify zero hidden neurons for a linear model.
Maximum Number of Generations Without Improvement in Genetic Training Strategy
You may set the maximum number of generations without improvement that the algorithm will train on. The number of generations may be set between 10 and 1000 (integers only). This will allow you to control the length of training time.
Advanced Genetic Features
You can train your model using different goals of genetic optimization when using the Genetic Training Strategy. These goals include:
Maximizing R-Squared
Minimizing average error
Maximizing correlation
Minimizing Mean Squared Error (MSE)
Minimizing Root Mean Squared Error (RMSE)
Maximizing a user-definable number within tolerance
You can also have the Genetic Training Strategy minimize the number of unpredictable patterns and also to have it favor tighter fitting during optimization.
II. Real estate models - location, location, location
The answer to the following question applies not just to real estate applications, but probably to most applications of NeuroShell Predictor and NeuroShell Classifier
involving geography:
Q. I am building a real estate application. As you know, location is important in real estate, and I would like to use longitude and latitude of the property as an input. If I
convert either of these readings to say decimal, e.g. 46.22353 and -78.34095, can I use them as variables in NeuroShell Classifier or Predictor.
A. You can use longitude and latitude but the neural net will probably not be sensitive to very many significant digits, so it will not be able to distinguish between nearby
clusters. If it does distinguish, use of those variables will not relate homes in one cluster on the east side of town to a cluster of the same type of neighborhood on the
west side of town. So we don't think longitude and latitude are a good idea. We think it will be much more effective to pick up location using one of the following
techniques:
1. Make a variable from 1 to 10 say that describes the type of neighborhood (1 = small clapboard house beside the railroad tracks, 10=wealthy executive neighborhood).
2. Maybe better than 1 to 10 is an input variable that gives the average real estate price in the immediate neighborhood. Or maybe this variable could be used WITH the
above.
3. Build separate models (nets) for each type of neighborhood. Maybe one for each type of home or neighborhood - expensive homes, medium homes, working class
homes. You'll probably want a separate model at least for each city or section of a major city. That way all models are comparing apples to apples. Of course, getting
enough data comes into play here. If you find that there aren't too many examples in some model, use the genetic method, which is better with sparse data.
---
National Chung Hsing University Department of Animal Science
National Taipei University of Business Department of Business Administration
MingChi University of Technology Department of Industrial Engineering & Management
Minghsin University of Science & Technology The Teaching Center of Natural Science
Aletheia University Department of International Trading
Water Resources Planning Branch, Water Resources Agency, Ministry of Economic Affairs Geo-lechnical Engineering Division
Ming-Chi Primary School
Fu Jen Catholic University Department of Business Administration
National Taipei University of Business Department of Business Administration
MingChi University of Technology Department of Industrial Engineering & Management
Minghsin University of Science & Technology The Teaching Center of Natural Science
Aletheia University Department of International Trading
Water Resources Planning Branch, Water Resources Agency, Ministry of Economic Affairs Geo-lechnical Engineering Division
Ming-Chi Primary School
Fu Jen Catholic University Department of Business Administration