"Learning by Teaching" has proven highly effective in enhancing students' learning, especially if it offers individualized, adaptive support. Currently, the key pedagogical decisions���what and when to teach���are typically tailored for each specific Learning by Teaching system and STEM problem-solving domain. This research aims to develop a generalizable reinforcement learning framework to dynamically determine the what, when, and why of adaptive pedagogical support in Learning by Teaching systems. This framework will enable existing STEM problem-solving systems to provide adaptive scaffolding and intelligent, contextualized feedback tailored to each student's interactions with a teachable agent. It will expand the system���s capabilities to support student agency, foster trust through explanations, and offer a broader scope of adaptive assistance.

We propose to develop a transformative technology in the form of teachable agent to amplify the effect of learning by teaching that we shall call a smart teachable agent, or smart TA for short. The smart TA asks students questions to justify their reasoning while solving equations. When student���s response could be elaborated, the smart TA further provides a follow up question to solicit a response that reflects a connection between procedural operations and conceptual justifications.

In this project, researchers will develop and evaluate an online game-like environment for middle school students to solve and learn algebra linear equations by teaching a simulated peer student. Learning by teaching is a promising style of instruction, with evidence supporting that when students engage in peer tutoring there is a benefit for both the tutee and tutor.

We propose to develop learning-engineering methods to efficiently build an effective online STEM learning environment, in the form of adaptive online courseware called CyberBook, to promote robust mathematics learning with understanding. The proposed CyberBook is a combination of traditional online courseware (that promotes conceptual understanding) and intelligent tutoring systems (that support guided learning-by-doing). We hypothesize that these two well-established technologies can be combined by identifying the shared latent learning constructs, i.e., skills and concepts to be learned. We further hypothesize that the resulting cyberlearning space will promote synergetic learning that, by definition, will fertilize the desired proficiency.