ECE 539 Course Project NEURAL NETWORK APPROACHES FOR

Motivations £ Nowadays, fuel economy becomes a great concern of the governments and drivers

Data Description £ Source: UCI Machine Learning Repository http: //archive. ics. uci. edu/ml/datasets/Auto+MPG £

Data Preparation £ 392 sets of data £ Correlation coefficients between I/O were calculated

Linear Regression 7 -way cross validation £ Training MSE = 11. 12 Tuning MSE

Multi Layer Perceptron £ MATLAB Neural Network Toolbox Used £ Learning algorithms: £ £

Multi Layer Perceptron Structure: 7 -12 -1 feedforward network £ Log-sigmoid function for hidden

Conclusions and Future Work £ MLP yields better performance than linear regression after fine

Slides: 9

Download presentation

ECE 539 Course Project NEURAL NETWORK APPROACHES FOR AUTOMOBILE MPG PREDICTION 12/14/2010 Xiaofei Sun University of Wisconsin-Madison

Motivations £ Nowadays, fuel economy becomes a great concern of the governments and drivers £ MPG varies with vehicle specs and conditions £ £ Database available online only accounts for different models £ Large amount of data required Build NN models to predict the MPG based on given specs and conditions £ MLP £ RBF 1/8

Data Description £ Source: UCI Machine Learning Repository http: //archive. ics. uci. edu/ml/datasets/Auto+MPG £ 8 Inputs: 1. cylinder # 2. displacement 3. horsepower 4. weight 5. acceleration 6. year 7. origin 8. manufacturer £ 1 Output: MPG 2/8

Data Preparation £ 392 sets of data £ Correlation coefficients between I/O were calculated 0. 9 0. 7 0. 6 0. 5 0. 4 0. 3 0. 2 0. 1 r M an uf ac O tu ri g re in r ea io at er cc el A Y n t gh ei W Po or se isp D H la c yl in em w er en t de r 0 C Correlation Coefficient 0. 8 3/8

Linear Regression 7 -way cross validation £ Training MSE = 11. 12 Tuning MSE = 12. 70 50 45 45 40 35 R 2 = 0. 8476 35 30 30 Target 40 R 2 = 0. 8335 25 20 15 15 10 10 5 5 0 0 0 10 20 Output 30 40 4/8

Multi Layer Perceptron £ MATLAB Neural Network Toolbox Used £ Learning algorithms: £ £ Gradient descent with momentum £ Scaled conjugate gradient £ Levenberg-Marquardt Datasets were randomly divided into three subsets: £ 60% for training £ 20% for validation (early stopping) £ 20% for testing 5/8

Multi Layer Perceptron Structure: 7 -12 -1 feedforward network £ Log-sigmoid function for hidden layer £ Linear function for output layer Test MSE = 5. 11 Training MSE = 4. 03 50 45 40 35 30 25 20 15 10 5 0 R 2 = 0. 9346 Target £ 0 10 20 30 Output 40 50 45 40 35 30 25 20 15 10 5 0 R 2 = 0. 9157 0 10 20 30 Output 40 50 6/8

Conclusions and Future Work £ MLP yields better performance than linear regression after fine tuning 14 12 10 8 MLP 6 Linear Regression 4 2 0 Training MSE £ Test MSE Will construct radial basis function network, and compare with MLP 7/8

Any Questions? ? 8/8