Learning from human demonstration
Brandon Packard, Santiago Ontañón · 2018
Learning from demonstration (LfD) is a branch of machine learning that focuses on learning how to perform a given task by observing a demonstrator perform one or several demonstrations of it. Moreover, many of the current LfD techniques assume a large pool of training examples from which to learn. The long term goal of our research is to develop general LfD methods which can more easily learn from human demonstrators than state-of-the-art methods, through requiring less training data and more carefully selecting when to obtain more training data. The main contribution of this thesis is a novel Active Learning from Demonstration algorithm called SALT (Selective Active Learning from Traces), an algorithm which can match or outperform other state-of-the-art algorithms in multiple domains, while using less training data. We have also gathered evidence that human demonstrators find it preferable to another state-of-the-art LfD algorithm, that it is less mentally burdensome for them to train, and that it learns better from them.