University of Twente Student Theses

Login

An investigation of generative replay in deep reinforcement learning

Imre, Baris (2021) An investigation of generative replay in deep reinforcement learning.

[img] PDF
1MB
Abstract:Catastrophic forgetting is an issue that persists in all deep reinforcement learning settings. Traditionally, catastrophic forgetting is solved by using experience replay, which is a buffer that remembers past experiences and uses these experiences to train an agent. This solution is effective in overcoming catastrophic forgetting but it heavily depends on memory. It would be a big breakthrough, if catastrophic forgetting could be solved without using memory, since this would dramatically increase scalability. To this end, a generative replacement is proposed in this paper for experience replay. Variational autoencoders are tested as a generative model on a simulation setup with a state of the art deep reinforcement learning algorithm. This method is called generative replay. Multiple methods of training and interaction are tested in order to explore the combination of generative replay and a deep reinforcement learning agent. Even though this paper does not present overwhelmingly positive results, it gives many insights on combining these networks in an experimental setup, and explores future possibilities for this approach.
Item Type:Essay (Bachelor)
Clients:
University of Twent
Faculty:EEMCS: Electrical Engineering, Mathematics and Computer Science
Subject:54 computer science
Programme:Computer Science BSc (56964)
Awards:Best Paper Award
Link to this item:http://purl.utwente.nl/essays/85772
Export this item as:BibTeX
EndNote
HTML Citation
Reference Manager

 

Repository Staff Only: item control page