Impact of Feature Representation for Imitation Learning in Autonomous Drive

Abstract

Autonomous drive in complex traffic scenarios is a demanding task to solve. With high-dimensional input data available, problems related to redundancy and irrelevance are often implicated, hence determining what features bring the most useful information is of vital importance. The purpose of this thesis is to investigate how different dimensionality reduction methods affect the driving performance and how to determine what features are most relevant. Specifically, these questions were studied in a simulated environment where a car is manoeuvred using deep neural networks through a sequence of signalised intersections. Four different dimensionality reduction methods have been studied: choice of features based on reason, Principal Component Analysis, Auto-Encoders and Integrated Encoders. The results showed that the models which used a feature representation based on reason were shown to perform best. Also, the weight distributions of a model using all available features indicated that influential features may be partially identified by studying the spread of the weights. Therefore, an approach is proposed where the choice of features should be based on reason as well as a study of the features’ respective set of weights. In conclusion, establishing the most relevant feature representation is important since it may benefit the training of the models.