Accelerated Learning and Potential Explainability Gains Through Object-Based Deep Reinforcement Learning

Jong, Niels de (2022) Accelerated Learning and Potential Explainability Gains Through Object-Based Deep Reinforcement Learning. Master's Thesis / Essay, Artificial Intelligence.

Preview	Text mAI_2022_deJongNA.pdf Download (1MB) | Preview
	Text toestemming.pdf Restricted to Registered users only Download (118kB)

Abstract

In the last two decades, artificial intelligence (AI) has seen considerable progress due to advances in deep learning (DL). However, the challenges of data-hungriness and model explainability seem to persist in DL's most popular branch, namely supervised DL. Hence, researchers have recently argued to shift attention towards non-supervised DL, such as deep reinforcement learning (DRL). Within DRL, a promising and relatively new avenue involves the use of high-level features ('objects') in models, giving rise to object-based DRL. Motivated by recent results, we investigate whether object-level state representations accelerate learning in DRL, and tentatively attempt to answer whether they make DRL methods more explainable. We do so by conducting three experiments in which we compare, against non-object baselines, the average undiscounted return curves of two existing and one proposed object-based DRL method, while also extracting object saliency maps (OSMs). Results indicate that learning may indeed be accelerated, provided that a sufficiently effective DRL method is chosen and that the representations are presented consistently. Further, using the collected OSMs, we present a critical discussion of the potential use of the OSM as an object-based explainability tool for DRL methods, and suggest how this possible use may be evaluated in future research.

Item Type:	Thesis (Master's Thesis / Essay)
Supervisor name:	Sabatelli, M.
Degree programme:	Artificial Intelligence
Thesis type:	Master's Thesis / Essay
Language:	English
Date Deposited:	15 Jul 2022 14:04
Last Modified:	15 Jul 2022 14:04
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/27927

Actions (login required)

View Item