It’s not Magic After All – Machine Learning in Snap! using Reinforcement Learning

The societal relevance of artificial intelligence is growing rapidly. Advances are primarily driven by machine learning techniques. Recently, many educational tools for teaching AI have been introduced, allowing the user to implement AI features within pedagogical programming environments. However, many of these existing approaches share a common trait: the implementation of the underlying AI framework remains a black box, where external API calls or servers handle the actual computing. For the user, this typically means there is no chance to “see inside” the implementation. As a result, users often cannot gain a deeper understanding of how the “computer is learning”. In this paper, we propose design principles for a framework in order to break open that black box. These design principles are implemented in the first part of SnAIp, a project aimed at enabling Machine Learning within Snap!. The focus of this part is using Reinforcement Learning within Snap! games. The corresponding framework enables constructionist learning and is implemented entirely in Snap!, which allows for a high degree of transparency and tangibility. Furthermore, we present a curriculum for Reinforcement learning using the SnAIp framework. With this, we outline a way to address ML in the classroom using block-based languages, while enabling the all-important “look behind the scenes”.