Dennis Konkol
Benchmarking of Deep Reinforcement Learning Algorithms on Embedded GPUs
Abstract
This bachelor’s thesis on Deep Reinforcement Learning (DRL) algorithms compares multiple
models on various embedded GPUs. The focus is on evaluating DRL algorithms’ performance
and power consumption, particularly DQN and DDPG. Base implementations in Python (stablebaselines3),
available as open source, serve as the starting point. Two specific implementations
are considered: one for engineered motor control from the University of Paderborn and a freely
chosen application from the OpenAI Gym environment. Here, I chose the CartPole environment
which will be discussed later in this thesis.
Furthermore, a custom DQN implementation for CUDA GPUs is developed, benchmarked, and
compared to the Python base implementation.