Efficient Preference-based Reinforcement Learning

• Main Author: Wirth, Christian
• Format: Others
• Language: en
• Published: 2017
• Online Access: https://tuprints.ulb.tu-darmstadt.de/6952/1/ThesisColorMerged.pdf; Wirth, Christian <http://tuprints.ulb.tu-darmstadt.de/view/person/Wirth=3AChristian=3A=3A.html> (2017): Efficient Preference-based Reinforcement Learning.Darmstadt, Technische Universität, [Ph.D. Thesis]

Common reinforcement learning algorithms assume access to a numeric feedback signal. The numeric feedback contains a high amount of information and can be maximized efficiently. However, the definition of a numeric feedback signal can be difficult in practise due to several limitations and badly defined values may lead to an unintended outcome. For humans, it is usually easier to define qualitative feedback signals than quantitative. Hence, we want to solve reinforcement learning problems with a qualitative signal, potentially capable of overcoming several of the limitations of numeric feedback. Preferences have several advantages over other qualitative settings, like ordinal feedback or advice. Preferences are scale-free and do not require assumptions over the optimal outcome. However, preferences are difficult to use for solving sequential decision problems, because it is unknown which decisions are responsible for the observed preference. Hence, we analyze different approaches for learning from preferences and show the design principles that can be used, as well as the advantages and problems that occur. We also survey the field of preference-based reinforcement learning and categorize the algorithms according to the design principles. Efficiency is of special interest in this setting, as it is important to keep the amount of required preferences low, because they depend on human evaluation. Hence, our focus is on efficient use of the preferences. It can be stated that it is important to be able to generalize the obtained preferences, as this keeps the amount of required preferences low. Therefore, we consider methods that are able to generalize the obtained preferences to models not yet evaluated. However, this introduces uncertain feedback and the exploration/exploitation problem already known from classical reinforcement learning has to be considered with the preferences in mind. We show how to efficiently solve this dual exploration problem by interleaving both tasks, in an undirected manner. We use undirected exploration methods, because they scale better to high-dimensional spaces. Furthermore, human feedback has to be assumed to be error-prone and we analyze the problems that arise when using human evaluation. We show that noise is the most substantial problem when dealing with human preferences and present a solution to this problem.