九十七學年度下學期 類神經網路 研究計畫書

 

一、研究計畫中英文摘要:

 

This research project aims at KDD Cup 2009 Customer Relationship Management provide to data to proceed a fast classification management and have already provided three parts of the target variable and a great deal of customer to train data in the contest of current, but it amount of data but greatly have to frighten, the general example program performance gets up to all rather take a lot of doing or make the computer crashed, so the research project research point how reach real quickly categorize.

 

Want to reach a fast classification and then have to find out good algorithm, according to currently collect of suggest in the data, the neural network is a kind of algorithm that can learn and remember by oneselfmost likely used to imagespeechpath to programdata analysis etc, so apply set of theory in this contest and definitely can reach to well recognize rate.

 

In the program design is make use of MATLAB mathematics tool software to carry on a fast calculation, because the neural network has already a history, so MATLAB has already had a sound letter type in program can provide a direct usage.

 

 

 

Key Wordneural networkMATLABKDD Cup 2009

 

 

 

二、研究計畫內容:

(一)研究計畫之背景及目的。

 

        KDD Cup is a world data processing contestwill hold in every yearthe content of topic includes many realmsmedical treatment, customer management for example, but in the data measure aspect usually is greatly have to frighten, the general computer equipment maybe can't compute, so sponsor conveniently of contest collect the world to the data processing have the interest to common research, how reach effectively and fast data processing with the decrease amount of calculation and suggest a upgrading efficiency way.

 

The things that aim at different realm all have it the characteristic to order, that wants to turn by data reaches the effect of estimate, usually the collection of data is necessary, but want to raise an accurate degree, the amount of data is opposite of become a direct proportion to go up for the meeting, in the process of operating the huge data certain will cause last equipment rather big burden, operation to definitely allow of not small, so the plan want to try to make use of mathematics software of a type of neural network and MATLAB to analyze big amount of data for research , how promote operation speed and decrease to operate quantity is this research point.

 

The type of predict to be more useful than weather forecast, the stock market analysis, speech recognition and image to recognize aspect through the data processing currently, if the ability is accurate to promote to a minimum error on the research, the people of will have more convenient in life, with have the suggestion that the more data turns on the application.

 

(二)        研究方法、進行步驟及執行進度

 

1.  Study method and reason

I will adoption to Back-propagation Neural Network structure to carry on a processing in this research, because the data provided by this contest belongs to big amount of data, and pour to deliver a type of nerve network characteristic in which is handle big amount of data, but in time need more time to handle, and the Hopfield Neural Network then handles time faster, but the amount of processing of the data has restriction, therefore I will adoption to deliver the way processing of a type of nerve network.

 

Can be divided in to pour to Back-propagation Neural Network and the Hopfield Neural Network in the type the nerve network, but they are each own strong point on the data processing, the following is both of merit and fault:

 

Back-propagation Neural Network

Merit

1.  Learning an accurate degree is high, can handle complicated sample to identify.

 

2. It's quick to remember speed.

 

Fault

1.  Learn speed slowly, refrain from action speed slowly, usually want several hundred or the thousands of learning circularly can 

refrain from rash action.

 

2.  Different beginning's to weighted value and biased weighted value will cause dissimilarity of refrain from rash action a result, generally speaking, it will be convergence action to a partial minimum error margin to as a result, we have to the beginning using many different weighted value and biased weighted value to look for the minimum error margin value of one reasonable, as for what refrain from rash action a result for the best, cannot know.

 

Hopfield Neural Network(HNN)

Merit

1.  This way is a kind of parallel input and parallel output's network structures, so has the high-speed operation ability of in great quantities parallel calculation.

 

Fault

1.  Because the network only convergence action to a partial minimum value that adjoin to beginning an appearance value, so different of the beginning start an appearance value will get different solution, we have to the beginning using many dissimilarities start an condition value to solve to look for a more reasonable solution.

 

2.  The enactment of good and bad with parameter answered very relevant, but the enactment of parameter don't have a method for systematizing to follow, can only take trying a false way as it.

 

2. Anticipate the difficulty of possible situation

This project may suffer to cause most in the process of designing because the data quantity is too huge computer processing up of delay, if the possibility needs to cost much time to wait for in every time the training the process, but each data for train not sure will complete smoothly, so single data for writing anticipates possibility need to train several times, waiting for of time certainly will become necessary.

 

(三)預期完成之工作項目及成果。

 

Expecting can complete all of the event, but on the efficiency take completion as a final target to carry on first and for the time being and in spite of.