Feature-directed Active Learning for Learning User PreferencesDownload PDF

Anonymous

Published: 24 May 2019, Last Modified: 05 May 2023XAIP 2019Readers: Everyone
Keywords: Preferences, Active Learning, Planning
TL;DR: Learning preferences over plan traces using active learning.
Abstract: Learning preferences of users over plan traces can be a challenging task given a large number of features and limited queries that we can ask a single user. Additionally, the preference function itself can be quite convoluted and non-linear. Our approach uses feature-directed active learning to gather the necessary information about plan trace preferences. This data is used to train a simple feedforward neural network to learn preferences over the sequential data. We evaluate the impact of active learning on the number of traces that are needed to train a model that is accurate and interpretable. This evaluation is done by comparing the aforementioned feedforward network to a more complex neural network model that uses LSTMs and is trained with a larger dataset without active learning.
9 Replies

Loading