A significant amount of work has advocated that Learning from Demonstration (LfD) is a promising approach to allow end-users to create behaviors for in-game characters without requiring programming. However, one major problem with this approach is that many LfD algorithms require large amounts of training data, and thus are not practical for learning from human demonstrators. In this paper, we focus on LfD with limited training data, and specifically on the problem of active LfD where the demonstrators are human. We present the results of a user study in comparing SALT, a new active LfD approach, versus a previous state-of-the-art Active LfD algorithm, showing that SALT significantly outperforms it when learning from a limited amount of data in the context of learning to play a puzzle video game.

A significant amount of work has advocated that Learning from Demonstration (LfD) is a promising approach to allow end-users to create behaviors for in-game characters without requiring programming. However, one major problem with this approach is that many LfD algorithms require large amounts of training data, and thus are not practical. In this paper, we focus on LfD with limited training data, and specifically on the problem of Active Learning from Demonstration in settings where the amount of data that can be queried from the demonstrator is limited by a predefined budget. We extend our novel Active Learning from Demonstration approach, SALT, and compare it to related LfD algorithms in both task performance (reward) and similarity to the demonstrator's behavior, when used with relatively small amounts of training data. We use Super Mario Bros. and two variations of the Thermometers puzzle game as our evaluation domains.

Learning from Demonstration has the potential to enable the crafting of behavior for non-player characters, allies, and enemies without requiring programming knowledge. This paper focuses on addressing two key problems of LfD when applied to games. The first is data sequentiality, when actions might be influenced by previous environmental states/actions, instead of just the current state. The second is having structured representations of data, where data is provided as an arbitrary number of predicates instead of a fixed-length vector. In this paper, we evaluate a collection of feature selection strategies to address these problems in the context case-based learning algorithms in the domain of Minecraft.