Introduction

Based on the lab tutorial in neural networks, the task is to implement your own estimation system using a neural network with Bayesian Regularization.

The dataset describes fuel consumption on bus routes, with several input parameters and a ground truth fuel consumption column. This is used to estimate the fuel consumption of a planned bus route, with some co-variance error. The case closely resembles the forest log pile case from the tutorial. You are allowed to set all the parameters yourself and to design the neural network as you see fit. You are also allowed to take the error assessment script from the tutorial.

Include the MATLAB script code, the training diagrams and the best- case co-variance error that you can achieve. You are free to implement the neural network as you see fit. You are allowed to use  the tutorial example as a template.

Some things to think about when you do the task:

1. Identity numbers in features

The identity numbers of the bus drivers, vehicles etc are very random. For a Support Vector Machine or any other non-nn method this will severely degrade the estimation accuracy when using the id of the drivers. There is also a degrading effect on the accuracy when using neural networks, although not as severe. In order to achieve  higher accuracy, you should think about a way to sort/reset id numbers so they correspond better to other features in the dataset.

One way would be to sort the id numbers according to the mean fuel consumption connected to that ID. The downside of that would be    that you would have to retrain the network when a new ID (new staff, new bus etc) is introduced into the dataset. Can you think of a  way to sort the data where constant retraining would not be necessary?

2. List of available features