Avoiding Brownian motion by utilizing consequential action selection in reinforcement learning for robotics

Brink, Ivo (2022) Avoiding Brownian motion by utilizing consequential action selection in reinforcement learning for robotics. Bachelor's Thesis, Artificial Intelligence.

Preview	Text Brink Consequential Action Thesis.pdf Download (1MB) | Preview
	Text Toestemming.pdf Restricted to Registered users only Download (127kB)

Abstract

Robotics solutions often require high maintenance and possess low adaptability in the workspace and are therefore not as widely adopted in the industry as envisioned. Reinforcement learning is believed to be able to alleviate the disadvantages mentioned above. Regardless, progress has been slow, and utilized performance benchmarks are simplified representations of realistic working conditions. Furthermore, trajectories of seemingly random small state changes are observed in exploration phases that seem redundant for learning. This research uses more versatile benchmarks to test a different approach to action selection in reinforcement learning for robotics. A heuristic is developed that selects more meaningful actions instead of the actions that result in the so-called Brownian motion trajectories. These consequential actions are efforts that are estimated to provide more significant state differences in the utility landscape, with an anticipated increase in exploration and faster rewards. A state-of-the-art Soft Actor-Critic algorithm has been developed to train an UR5e robot in three different tasks. Results provided evidence of the existence of Brownian motion in the initial training phases over all three tasks. Selecting consequential actions provided more effective exploration in the early phases and, as a result, achieve an average increase in success rate of 10%. Consequential action selection proved less effective in later stages of learning, causing model convergence

Item Type:	Thesis (Bachelor's Thesis)
Supervisor name:	Schomaker, L.R.B. and Luo, S.
Degree programme:	Artificial Intelligence
Thesis type:	Bachelor's Thesis
Language:	English
Date Deposited:	20 Dec 2022 15:11
Last Modified:	20 Dec 2022 15:11
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/29059

Actions (login required)

View Item