Reference Type | Thesis |
Author(s) | Tschirner, J. |
Year | 2018 |
Title | Boosted Deep Q-Network |
Journal/Conference/Book Title | Bachelor Thesis |
Abstract | Deep reinforcement learning achieved outstanding results in the last years, since it has emerged from reinforcement learning combined with the capabilities of deep neural networks. This thesis is about Boosted Deep Q-Network (BDQN), a novel deep reinforcement learning algorithm that employs gradient boosting to estimate the action-value function in reinforcement learning problems. The aim with this approach is to meet some of the major challenges still present in the field. On the one hand, they consist of the necessary model complexity for high-dimensional problems and moreover, of the requirement to make efficiently use of the environment’s data. BDQN was evaluated empirically first on two standard control problems, and then on one more complex environment with a high-dimensional observation space. In comparison to Deep Q-Network (DQN), the algorithm performed competitively on the standard tasks using a smaller model complexity, but had serious problems learning a reasonable policy in the complex environment. |
URL(s) | /uploads/Team/SamueleTosatto/bachelor_thesis_jeremy_tschirner.pdf |
Link to PDF | /uploads/Team/SamueleTosatto/bachelor_thesis_jeremy_tschirner.pdf |
Language | English |