2022年神经网络原理英文 .pdf

《2022年神经网络原理英文 .pdf》由会员分享，可在线阅读，更多相关《2022年神经网络原理英文 .pdf（5页珍藏版）》请在淘文阁 - 分享文档赚钱的网站上搜索。

1、Artificial Neural NetworkPortion of Coil StudyLTC William M. CrocollSchool of Systems and Acquisition ManagementUS ARMY LOGISTICS MANAGEMENT COLLEGEBld 12500, 2401 Quarters Road, FT Lee, VA 23801-1705email: crocollwlee.army.mil, DSN 539-4231; Comm (804) 765-4231http:/www.almc.army.milINTRODUCTION TO

2、 ARTIFICIAL NEURAL NETWORKSA neural network is a massively parallel system comprised of many highlyinterconnected, interacting processing elements (also called nodes or neurons) based onneurobiological models of the brain (Dayhoff, 1990). A major task for a neural network is tolearn and maintain a m

3、odel of the world consistent enough to achieve the goals of the applicationof interest (Haykin, 1994). These systems act as non-linear, non-parametric function estimatorsthat learn to map inputs to outputs on a non-linear, multidimensional surface to fit general non-linear, multivariate functions (Z

4、urada, 1992).Neural networks exhibit many advantageous properties for solving complex problems.The following characteristics of neural networks emphasize their pattern recognition capabilities,making them particularly attractive for solving complex, data rich problems (Sheppard & Gent,1991). They:(1

5、) can learn from examples and adapt to new situations;(2) are robust (i.e., can abstract information from noisy, incomplete, and inaccurate data);(3) can generalize from examples (i.e., can provide correct solutions from data similar to but notexactly like training data);(4) can construct solutions

6、quickly with no reliance on domain knowledge;(5) can approximate any complex (ill-defined or unknown) multivariate function and form aclassification decision from the recognition of discriminating patterns;(6) are computationally efficient (i.e., have the speed to operate in real-time); and(7) can i

7、mplicitly account for the relative importance of input sources.名师资料总结 - - -精品资料欢迎下载 - - - - - - - - - - - - - - - - - - 名师精心整理 - - - - - - - 第 1 页，共 5 页 - - - - - - - - - Since neural networks are adaptable (i.e., learn from experience) there is no need for an apriori mathematical model for input-ou

8、tput transformation. They learn correct responses byobserving real world examples. With sufficient examples, neural networks exhibit stronggeneralization capabilities. Generalization is defined as the capability of producing reasonableresponses to inputs similar to, but not exactly like, the previou

9、s inputs encountered duringtraining. This adaptability also facilitates relatively easy retraining to deal with minor changes inthe environment being modeled (Haykin, 1994). Neural networks are not programmed buttrained. Because of this characteristic, they are best applied to problems whose solutio

10、n requiresknowledge which is difficult to specify but for which there is an abundance of examples(Sheppard & Gent, 1991). This makes neural networks ideal for data rich, theory poorapplications. They learn by adjusting the interconnection strengths between the artificial neuronsusing computationally

11、 efficient algorithms for discovering appropriate non-lineartransformations between the problem space and the solution space (Sheppard & Gent, 1991).The objective of learning is to determine the optimal interconnection strengths (i.e., weights) thatprovide the best approximation of the desired conti

12、nuous, multivariable function based on a setof training examples (Zurada, 1992). A properly trained network should then be able to respondcorrectly to input not previously seen. This would indicate that the neural network hasgeneralized the knowledge of the domain and not merely memorized the traini

13、ng data.CASCADE CORRELATION NETWORK OVERVIEWThe Cascade Correlation network (CCN) (Fahlman and Lebiere, 1990) is an example ofthe network growing approach. The procedure begins with a minimal network consisting of theappropriate application specified input nodes and output node(s) but no initial hid

14、den neurons.Hidden neurons are added to the network one by one resulting in a multilayered structure. Eachnew hidden neuron receives a connection from each of the input nodes and each pre-existinghidden neuron. Thus the structure is a deep net of cascaded hidden nodes as opposed to anetwork with a w

15、ide hidden layer. As each new hidden neuron is added, its input weights aretrained first. Then all of the weight connections to the output nodes are trained while leaving theother weights unchanged. Weights are adjusted to reduce overall network error typically usingeither the LMS rule or the Quickp

16、rop learning algorithm. This process continues untilperformance is judged adequate.名师资料总结 - - -精品资料欢迎下载 - - - - - - - - - - - - - - - - - - 名师精心整理 - - - - - - - 第 2 页，共 5 页 - - - - - - - - - WHY ARTIFICIAL NEURAL NETWORKS ARE APPROPRIATE FOR FORECASTINGCARAVAN INSURANCE POLICY PURCHASING.The pattern

17、 recognition capabilities of artificial neural networks (e.g., ability to learnfrom examples, approximate complex non-linear multivariate functions, handle noisy,incomplete, and inaccurate data, and generalize) lend themselves nicely to the characteristics ofthis insurance policy purchasing problem.

18、 This appears to be a complex, data rich, theory poortype of problem with substantial historical data depicting the characteristics of each person aswell as whether that person purchased a Caravan insurance policy or not. However, there is littletheory to rely upon to develop a specific functional (

19、read mathematical) form relating a person scharacteristics (input variables) to whether or not they purchased a Caravan policy. It can beassumed that some complex non-linear, multivariate function maps the input variables to policypurchase or not, thus this becomes a pattern recognition task. The ta

20、sk is to recognize the patternor relationship between the input variables and policy purchase. An artificial neural network willaccomplish that pattern recognition by approximating the unknown and unknowable functionmapping inputs to output and use that function approximation to make future predicti

21、ons about aperson s propensity to purchase a Caravan insurance policy.MODEL DEVELOPMENTThe initial task was to define the problem and objectives as well as select the variables toconsider in the model. In this study, the problem was defined by the CoIL Challenge contestspecifications. Likewise, CoIL

22、 provided the historical data. The data was used as provided andno further data transformations were conducted. The data was then divided into segments. The5822 records in the ticdata2000 was divided into two sets. One set consisting of 4322 records(278 who purchased a Caravan policy and 4044 who di

23、d not) was used for model development.The other set consisted of 1500 records (70 who purchased a Caravan policy and 1430 who didnot) and was used to test the neural network models. Various Cascade Correlation neuralnetwork models were investigated. All neural network models were developed and teste

24、d on a400 Mhz Compaq Desk Pro with Windows NT. A commercial software package (NeuroShellClassifier from Ward Systems Group) was used to model and test the neural networks.名师资料总结 - - -精品资料欢迎下载 - - - - - - - - - - - - - - - - - - 名师精心整理 - - - - - - - 第 3 页，共 5 页 - - - - - - - - - NeuroShell Classifier

25、 implements an algorithm called TurboProp 2 which is a proprietary variantof the cascade correlation algorithm.MODEL PERFORMANCE RESULTS (AKA. MODEL VALIDATION)A split sample validation technique was used to evaluate each neural network model sperformance. As mentioned previously, the data was divid

26、ed into a model development segmentand test segment. Once each neural network was trained on the model development set, it wasused to provide predictions on the test set. Since the correct answers to the test set were known,each neural network model s performance was evaluated. The best performing n

27、etwork utilized12 input variables and produced an accuracy of 71.43% for those, in the test set, that actuallypurchased the Caravan insurance policy. A classification and regression tree technique was usedfor the variable reduction.CONCLUSIONThe major conclusion that can be drawn from this study is

28、that, given the input variablesprovided, an artificial neural network is a feasible technique for forecasting a person s propensityto purchase a Caravan policy. This can be used for marketing decision making. In this limitedinvestigative proof of concept study, the Cascade Correlation network, repre

29、sented as aproprietary variant called TurboProp 2 in NeuroShell Classifier, produced satisfactory results.FORECASTS (IMPLEMENTATION)The actual implementation of a neural network for future forecasts is relatively straightforward. I simply presented the input variable values for the 4000 records in t

30、he prediction set(ticeval2000) to the best network model and the network produced the forecast as its output. Theforecast was the probability that an insurance policy would be purchased. I then rank ordered the4000 records in the prediction set and selected the top 800 as those most likely to purcha

31、se aninsurance policy. The only drawback to the use of a neural network for this application is thelack of an explanation for the forecast. The forecasts are a function of the network weightswhich were developed during the training process. These weights, collectively, represent the名师资料总结 - - -精品资料欢

32、迎下载 - - - - - - - - - - - - - - - - - - 名师精心整理 - - - - - - - 第 4 页，共 5 页 - - - - - - - - - knowledge of the relationship between the input and output variables but is nearly impossible totranslate into a semantic representation.REFERENCESBeastall, W. D. (1989). Recognition of radar signals by neural

33、 networks. In Proceedings of theIEE Conference on Artificial Neural Networks, 139-142.Dayhoff, J. E. (1990). Neural Network Architectures: An Introduction. New York: Van Nostrand Reinhold.Fahlman, S. E. and Lebiere, C. (1990). The cascade-correlation learning architecture. In D.S.Touretzky (Ed.), Ad

34、vances in Neural Information Processing Systems 2 (pp. 524-53).San Mateo, CA: Morgan Kaufmann.Fredrickson, S. E., & Tarassenko, L. (1995). Text-independent speaker recognition using neuralnetwork techniques. In Proceedings of the IEE 4th International Conference on ArtificialNeural Networks, 13-18.H

35、aykin, S. (1994). Neural Networks: A Comprehensive Foundation . New York: MacMillanPublishing Company.Leung, H., Blanchette, M., & Haykin, S. (1993). Clutter cancellation and sea ice detection usingartificial neural networks. In Proceedings of the SPIE-The International Society forOptical Engineerin

36、g: Applications of Artificial Neural Networks IV , 1965, 312-321.Moody, J., & Darken, C. J. (1989). Fast learning in networks of locally-tuned processing units.Neural Computation, 1, 281-294.Paez, T. L. (1993). Neural networks in mechanical system simulation, identification, and assessment. Shock an

37、d Vibration, 1(2), 177-199.Sheppard, C. P., & Gent, C. R. (1991). A neural network based sonar classification system. In Proceedings of the Europe 1991 MILCOMP Conference: Military Computers- Systems and Software, 181-186.SPSS, Inc. and Recognition Systems Inc. (1998). Neural Connection 2.1. Chicago, IL.Ward Systems Group, Inc. (1998). NeuroShell Easy Classifier. Frederick, MD.Zurada, J. M. (1992). Introduction to Artificial Neural Systems. New York: West PublishingCompany.名师资料总结 - - -精品资料欢迎下载 - - - - - - - - - - - - - - - - - - 名师精心整理 - - - - - - - 第 5 页，共 5 页 - - - - - - - - -