Transfer of Experience Replay in Deep Reinforcement Learning

Mueller, Luca (2022) Transfer of Experience Replay in Deep Reinforcement Learning. Bachelor's Thesis, Artificial Intelligence.

Preview	Text Thesis_Luca_Mueller_s3745503.pdf Download (1MB) | Preview
	Text toestemming.pdf Restricted to Registered users only Download (95kB)

Abstract

Transfer Learning (TL) is an area of machine learning concerned with adjusting existing models to fit new data. Deep Reinforcement Learning (DRL) combines Reinforcement Learning (RL) and Deep Learning to solve games and real world problems that can be modeled as Markov Decision Processes (MDP). Since training of DRL algorithms is expensive and often not possible due to limited availability of data, this area would greatly benefit from TL. Attempts to transfer model parameters have not been successful. A possible solution is transferring trajectories stored in an experience replay buffer to provide training data from a different task, provided that both source and target task are similar in nature. Two environments from the OpenAI gym were chosen in combination with three DRL algorithms. Agents were trained on the provided environments as well as a slightly modified version of each environment. Transfer was tested from standard to modified environments and vice versa. For reference a transfer of model parameters and a transfer of both experience replay buffer and model parameters was included. The results show an improvement in one of the tasks, regardless of transfer method, and no significant difference in the second.

Item Type:	Thesis (Bachelor's Thesis)
Supervisor name:	Sabatelli, M.
Degree programme:	Artificial Intelligence
Thesis type:	Bachelor's Thesis
Language:	English
Date Deposited:	07 Mar 2022 09:57
Last Modified:	07 Mar 2022 09:57
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/26664

Actions (login required)

View Item