For this technical project, in the first section, extensive research was carried out to find the appropriate design for building Fuzzy Logic Controller. Efforts were made to come up with the best possible technical design which can be generalized and implemented in any car. Coming across several websites and video tutorials, it was realised that a generalized automatic braking system consists of Brake Pedal, Vacuum Booster, Piston, Master Cylinder, Fluid Lines, and Calipers/Brake Pads. Braking Mechanism was later learnt to understand the entire braking system in terms of processes.

 

Fuzzy Logic Controller was designed in Matlab software using Gaussian and Trim functions for obstacles and speed handling. Later rules were defined to simulate control of piston, master cylinder through the input from braking pads.

 

Second section was regarding the designing of an artificial neural network architecture to forecast sales of the company based on the history of the company sales in three (3) years given in the table. Given data included the monthly sales of the company, number of public holidays occurs in each month, school holiday and economy condition.  It was required to analyse the input data, in order to design and develop a prediction model based on back-propagation training algorithm to forecast company monthly sales for the year 2013.

 

During the research on Neural Network, appropriate tools and techniques were investigated to come up with best possible solution for creating a neutral network for sales predication. It was learnt that the Neural Network Toolbox in Matlab provides better solution to create a neural network.

 

Neural Network toolbox consists four types of tools:

 

1.    Fitting Tool

2.    Pattern Recognition Tool

3.    Clustering Tool

4.    Time Series Tool

 

Further research was required to choose appropriate Neuron Model (Transfer Function) for the development. Feed-forward Network architecture was the most appropriate to create a neural network as it fully fits with the purpose.

 

After identifying appropriate tools for managing data set i.e. Fitting Tool, creating neural network architecture i.e. Feed-forward Network), Transfer functions for getting the output i.e. log-sigmoid and purelin. It is important to know which back-propagation training algorithms are most appropriate to train the network for improved performance and efficiency. For this, several Back-Propagation training algorithms were researched

 

1.    Levenberg-Marquardt (trainlm)

2.    Scaled Conjugate Gradient (trainscg)

3.    Resilient Back-Propagation (trainrp)

4.    BFGS Quasi-Newton (trainbfg)

 

During further investigation, it was learnt that the fastest training function among them all is Levenberg-Marquardt (trainlm) and is used to update weight and bias values according to Levenberg-Marquardt optimization. Scaled Conjugate Gradient (trainscg) and Resilient Back-Propagation (trainrp) were not considered as suitable choices because they are used for training large networks, mostly for pattern recognition which is not the case.

 

Third and final section required creation of a genetic algorithm for a newspaper organisation that needed to deliver newspapers all over the country.  Since on-time delivery service plays an important role in the organisation’s operations, therefore, the company needed to optimise its delivery service in order to cover all the cities in the shortest path.  Information such as the name of cities and the distance between them was already given and therefore, it was required to create and use genetic algorithm to optimise the newspaper delivery service of the company.