Design of on-demand traffic guidance systems

Abstract

The modern day car is equipped with a GPS navigation system which allows users to

predict their time of arrival at a destination when the recommended route is followed.

However, living in a highly urbanized city brings forth the inevitable problem of

congestion which undermines the effectiveness of GPS. What is more frustrating than

being stuck in traffic is knowing that it could have been avoided it if one’s GPS navigation

system had not brought him/her through this patch of bad traffic. Modern day GPS

navigation devices are sophisticated devices but lack the ability to predict prevailing

traffic conditions solely relying on the shortest path algorithm to navigate users to their

destination.

In this report, we investigate methods to create low-dimensionality models from highdimensionality

models by applying Principal Component Analysis and Kernel Principal

Component Analysis to different sets of data in hopes of creating low-dimensional models

of training data by scaling the dimensionality of the datasets. The resultant models can

then be effectively used in machine learning to create an algorithm which will allow small

devices which are low in computational power to recommend routes based on prevailing

traffic conditions.

The results show that Principal Component Analysis highly effective in creating a lowdimensional

model using just 33.2% of the total number of components with a 5% margin

of error. Kernel Principal Component Analysis allowed a dimensionality reduction of 39%

on a simulated dataset but is non-conclusive that the same percentage can be achieved

when applied on a high-dimensional model due to the complexity of mercer kernels and

uncertainty of the existence of the pre-image to convert a projection of a vector into feature

space back into input space.

The findings of this report is part of an effort to create a system which will be able to

predict traffic based on past statistical data. The findings of this report will contribute to

future research on Principal Component Analysis and Kernel Principal Analysis on highdimensional

models.