Effective reinforcement learning through intrinsic motivation and visual external memory in partially observable environments

Date

2023-9

Author

Demirbilek, Burak Han

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

218
views

146
downloads

Cite This

Reinforcement learning in practical scenarios often includes partial observability that requires long-term remembering of visual observations to obtain optimal policies. Addressing this challenge, this study introduces agents augmented with visual external memories, enhancing agents decision-making capabilities by constructing a context derived from both current observations and memory data. Moreover, to ensure effective utilization of the external memory for the agent, intrinsic motivation is incorporated as a secondary reward system, promoting long-term beneficial behaviors of using memory. Key contributions from this study include a novel framework for integrating visual external memory in reinforcement learning agents, the development of intrinsic motivation functions to efficiently learn how to utilize external memory to improve overall learning, empirical evaluations and experiments in various environments, and detailed comparison and analysis against the state-of-the-art. The results highlight the potential and advantages of the proposed approaches and present numerous possibilities for future investigation within this particular field of study.

Subject Keywords

Reinforcement learning, Partial observability, External memory management, Intrinsic motivation

URI

https://hdl.handle.net/11511/105568

Collections

Graduate School of Natural and Applied Sciences, Thesis