Folie 1

1
Introduction

• An ANN based approach to estimation of car fuel
  consumption
• Multi Layer Perceptron as a choice of the art
• Different discrete and continuous features
• Prediction without extensive trials on the road
• Could be used as first guess

2
Data Description

• Seven features used Horsepower, Weight, Number
  of Cylinders etc.
• Some data with unknown values
• Wide variation of cars, 398 samples available
• Equally distributed from different manufacturers
• Quite old, from early 80s

3
Data Preprocessing

• Feature dimension statistical values differ
  very much
• Removing name of car and values with unknown
  parameters
• Normalization of each feature dimension
• Create data sets for training and final testing

4
MLP Development I

• Estimation of underlying physical model
  difficult
• How complex and how much nonlinear?
• One hidden layer versus several hidden layers
• Cross-validation as best way to find optimal
  configurations
• Five parameters for variation

5
MLP Development II

• Each cross validation performed several times
  with same parameters to get a meaningful average
• Learning rate and epoch size most important
• Tables to evaluate best settings
• No complete automation, two-step evaluation
• Several comparable best configurations

6
Comparison to Base Case

• Best result from final training still 30-40
  worse than base case
• No improvement achieved with two hidden layers
• Results still good for first estimation
• Understanding of model not sufficient enough

7
Conclusion

• Use of MLP delivered satisfactory results, but
  not better than base case
• Using different activation function could bring
  improvement
• Without some prior knowledge of physical model
  hard to see what features are more important than
  others
• Other ANN like radial basis functions possible