Dimensionality Reduction Methods Using VAE for Deep Reinforcement Learning of Autonomous Driving
Yuta Uehara, Susumu Matumae · 2023
In recent years, significant progress has been made in the field of Deep Reinforcement Learning. However, the advancement of Deep Reinforcement Learning comes at the cost of requiring substantial amounts of data, which increases the learning cost as the process complexity grows. This challenge is particularly pronounced in domains such as autonomous driving, where a multitude of input data, such as camera images, sensor readings, and GPS coordinates, needs to be processed. Such data is inherently high-dimensional and necessitates considerable computational resources for concurrent processing and learning. In this paper, we investigate the impact of employing dimensionality reduction techniques, applied to images, on the learning process of autonomous driving. By implementing a Variational Autoencoder (VAE) to reduce the dimensions of input data while retaining learning accuracy, we aim to mitigate the learning costs. Our simulation results demonstrate that using VAE for dimensionality reduction yields performance equal to or surpasses that achieved when employing a naive pooling technique or directly using image data. Furthermore, our approach resulted in a 30% reduction in memory usage and a 50% decrease in training time.