Você está na página 1de 31

A Revolutionary Method to do

Manufacturing Demand Forecasting


with SVM - Support Vector Machines
Dr. Thomas Wölfl, TM3 Software
© 2010 Supply Chain Council. ALL RIGHTS RESERVED. | < filename > | Slide 1 | 25 October 2010
Agenda

1. Demand planning

2. Methods for demand forecasting

3. Artificial intelligence and machine learning

4. Support vector machines (SVM)

5. Example and results

© 2010 Supply Chain Council. ALL RIGHTS RESERVED. | < filename > | Slide 2 | 25 October 2010
Demand Planning

• Improve the accuracy of forecasts


• Goal: Improve production scheduling or inventory stocks

Manufacturer Party 1 Party 2


End-Customer
[Carbonneau et al., 2007]

• Each party in the supply chain does some processing on


the demand signal

© 2010 Supply Chain Council. ALL RIGHTS RESERVED. | < filename > | Slide 3 | 25 October 2010
Demand Planning

• Forecasting based on past demand

• Data Mining:
• Use of statistical tools
• Experience of market experts
© 2010 Supply Chain Council. ALL RIGHTS RESERVED. | < filename > | Slide 4 | 25 October 2010
Demand Planning

• Forecasting based on partner forecasts

• Combine forecasts:
• Use of data tools
• Experience of market experts
© 2010 Supply Chain Council. ALL RIGHTS RESERVED. | < filename > | Slide 5 | 25 October 2010
Demand Planning

• Combine both ways:


• Use past demands
• Use customer forecasts

• Add external information influencing demand


• Market figures, seasonal factors, …

• Learn the experience of maket experts

Use Artifical Intelligence (AI) and


Machine Learning (ML) technology

© 2010 Supply Chain Council. ALL RIGHTS RESERVED. | < filename > | Slide 6 | 25 October 2010
Agenda

1. Demand planning

2. Methods for demand forecasting

3. Artificial intelligence and machine learning

4. Support vector machines (SVM)

5. Example and results

© 2010 Supply Chain Council. ALL RIGHTS RESERVED. | < filename > | Slide 7 | 25 October 2010
Methods for demand planning

• Methods based on past demands:

• Naive Forecast
• Average
• Auto Regressive Moving Average (ARMA)
• Multiple Linear Regression
• …

© 2010 Supply Chain Council. ALL RIGHTS RESERVED. | < filename > | Slide 8 | 25 October 2010
Forecasting Methods

• Example: Demand of a certain article during the last year

Real demand

© 2010 Supply Chain Council. ALL RIGHTS RESERVED. | < filename > | Slide 9 | 25 October 2010
Forecasting Methods

• Demand forecast using


Auto Regressive Moving Average (ARMA)

Real demand Forecast for the next period: 75 pieces


ARMA forecast

• Problem: ARMA parameter selection


© 2010 Supply Chain Council. ALL RIGHTS RESERVED. | < filename > | Slide 10 | 25 October 2010
Forecasting Methods

• Demand forecast using


Modified Auto Regressive Moving Average (ModARMA)

Real demand Forecast for the next period: 54 pieces


ModARMA forecast

© 2010 Supply Chain Council. ALL RIGHTS RESERVED. | < filename > | Slide 11 | 25 October 2010
Agenda

1. Demand planning

2. Methods for demand forecasting

3. Artificial intelligence and machine learning

4. Support vector machines (SVM)

5. Example and results

© 2010 Supply Chain Council. ALL RIGHTS RESERVED. | < filename > | Slide 12 | 25 October 2010
Machine Learning

• Discipline of Artificial Intelligence (AI)

• Machine Learning (ML): An algorithm automatically


learns in order to make accurate predictions based on
past observations (Rob Schapire)

• Applications:
› Face detection: find faces in images
› Disease diagnosis: detect cancer in blood count
› Spam filtering
› Fraud detection
› Weather prediction
› …

© 2010 Supply Chain Council. ALL RIGHTS RESERVED. | < filename > | Slide 13 | 25 October 2010
Machine Learning

• Image classification example: Training

Class 1

Class 2

Class 3
Features of
training
Class 4 data

Class 5 Untrained
ML-Algorithm
Class 6
Given: training images and their categories
[Verbeek, 2009]

© 2010 Supply Chain Council. ALL RIGHTS RESERVED. | < filename > | Slide 14 | 25 October 2010
Machine Learning

• Image classification example: Testing

Class 3

Trained Classification
ML-Algorithm result

What are the categories


of these test images?

© 2010 Supply Chain Council. ALL RIGHTS RESERVED. | < filename > | Slide 15 | 25 October 2010
Machine Learning

• Image classification example: Testing

Class 4

Trained Classification
ML-Algorithm result

What are the categories


of these test images?

© 2010 Supply Chain Council. ALL RIGHTS RESERVED. | < filename > | Slide 16 | 25 October 2010
Machine Learning

• ML-Regression: Similar to classification, but the output


has the form of one or more real numbers
• Learns a function (e.g. the demand function)

• Example: predict gas mileage of a car based on number


of cylinders or doors, weight, horsepower, year, etc.

[Verbeek, 2009]

© 2010 Supply Chain Council. ALL RIGHTS RESERVED. | < filename > | Slide 17 | 25 October 2010
Machine Learning

• Machine Learning methods:


› Decision Trees
› Bayesian Networks
› Neuronal Networks
› …

› Support Vector Machines

© 2010 Supply Chain Council. ALL RIGHTS RESERVED. | < filename > | Slide 18 | 25 October 2010
Agenda

1. Demand planning

2. Methods for demand forecasting

3. Artificial intelligence and machine learning

4. Support vector machines (SVM)

5. Example and results

© 2010 Supply Chain Council. ALL RIGHTS RESERVED. | < filename > | Slide 19 | 25 October 2010
Support Vector Machines

• SVM-Demo: Credit ranking classification

• Input Variables:
› Net income of a person
› Living space / area of flat

• Classification:
› Good credit score
› Middle credit score
› Bad credit score

-t 2 -c 100 -g 100

© 2010 Supply Chain Council. ALL RIGHTS RESERVED. | < filename > | Slide 20 | 25 October 2010
Support Vector Machines

• SVM-Demo: Regression for demand forecasting

• Input Variables:
› Average demand of last period
› Forecast of this period
› Seasonal factor
› …

• Regression:
› Learn the demand function

-s 3 -t 2 -c 10 -g 1000 -p 0

© 2010 Supply Chain Council. ALL RIGHTS RESERVED. | < filename > | Slide 21 | 25 October 2010
Support Vector Machine

• SVM concept: Learn a separating „borderline“ between


the given data
Good credit score
Bad credit score
Income

Living Space
© 2010 Supply Chain Council. ALL RIGHTS RESERVED. | < filename > | Slide 22 | 25 October 2010
Support Vector Machine

• Which borderline to use?

Good credit score


Bad credit score
Income

Living Space
© 2010 Supply Chain Council. ALL RIGHTS RESERVED. | < filename > | Slide 23 | 25 October 2010
Support Vector Machine

• Select the borderline which maximizes the marign!

Good credit score


Bad credit score
Income

Margin
Width B
Margin
Width A
Living Space
© 2010 Supply Chain Council. ALL RIGHTS RESERVED. | < filename > | Slide 24 | 25 October 2010
Support Vector Machine

• What to do now? Good credit score


Bad credit score

Income

Living Space
© 2010 Supply Chain Council. ALL RIGHTS RESERVED. | < filename > | Slide 25 | 25 October 2010
Support Vector Machine

• Use a non-linear separation function


• Add a new feature: living space, income, age

Age

Income

Income

Living Space Living Space

© 2010 Supply Chain Council. ALL RIGHTS RESERVED. | < filename > | Slide 26 | 25 October 2010
Agenda

1. Demand planning

2. Methods for demand forecasting

3. Artificial intelligence and machine learning

4. Support vector machines (SVM)

5. Example and results

© 2010 Supply Chain Council. ALL RIGHTS RESERVED. | < filename > | Slide 27 | 25 October 2010
SVM demand forecasting
Car sales forecast

Real sales Traditional ARMA SVM forecast


figures forecast

[Wu, 2009]

© 2010 Supply Chain Council. ALL RIGHTS RESERVED. | < filename > | Slide 28 | 25 October 2010
SVM demand forecasting – key facts

• SVM is a field-proven method, e.g. used in


› medical informatics
› it-security

• SVM filters outliers and noise in demand data

• SVM handles multiple dimensions (features), large amount of


input data and non-linear relations

• Experiments show that SVM has a better forecast quality


(MAPE) then traditional methods

• SVM forecast quality is easy to measure: compare forecast


with historical data

© 2010 Supply Chain Council. ALL RIGHTS RESERVED. | < filename > | Slide 29 | 25 October 2010
Thomas.Woelfl@tm3-software.com

DEMAND FORECASTING WITH SVM -


SUPPORT VECTOR MACHINES

© 2010 Supply Chain Council. ALL RIGHTS RESERVED. | < filename > | Slide 30 | 25 October 2010
Support Vector Machine

• SVM-Regression: Learn the demand function

Demand

Time

© 2010 Supply Chain Council. ALL RIGHTS RESERVED. | < filename > | Slide 31 | 25 October 2010

Você também pode gostar