HCII Ph.D. Defense: Danny Weitekamp

Erik Harpstead - HCII, CMU
Kurt VanLehn - Arizona State University


Abstract
Interactive task learning (ITL) is a machine learning paradigm that envisions AI that can learn whole programs directly from non-programmers' natural instruction. In this dissertation I present a system called AI2T that improves upon an ITL sub-paradigm called authoring-by-tutoring, whereby highly adaptive educational technology known as intelligent tutoring systems (ITSs) are authored by teaching an agent with rapid human-like learning capabilities. In the course of about 20-30 minutes authors can tutor AI2T with demonstrations and interactive feedback instead of needing to program an ITS by hand; a process which typically requires 200-300 developer hours per hour of instruction.

Authoring-by-tutoring presents a significant opportunity to democratize the authoring of ITSs. The defining characteristic of an ITS is the automatic delivery of detailed step-by-step feedback and hints characteristic of human-to-human tutoring. ITSs are typically more effective than traditional instruction and in some cases even more effective than human tutors. Authoring-by-tutoring is a path toward building the cognitively focused, precisely engineered, and reliably accurate behaviors of traditional ITSs without needing to hand-program behaviors or rely upon costly pretrained AI systems like large language models (LLMs) that are prone to hallucinating incorrect solutions and feedback. Toward this aim, this work innovates on methods of machine-learning that robustly learns complex behaviors via rapid bottom-up induction, instead of by mimicking patterns in big-data.

In this dissertation, I present two novel machine-learning algorithms that enable a data-efficient and robust interactive task learning, whereby 100\% complete and accurate rule-based programs can be induced from interactive instruction. First I present STAND, a highly data-efficient algorithm for inducing preconditions for rules from binary reward signals. STAND out-performs algorithms like random forests and XGBoost known for their data-efficient learning on tabular data. STAND also enables a measure called instance certainty, an estimate of prediction probability that is more highly correlated with actual increases in holdout set performance than methods that rely on weighted ensembles. I show in simulation and with users that instance certainty can help authors estimate when AI2T has induced 100% complete programs, and show that it can provide active-learning support, helping authors identify the most helpful problems to tutor AI2T on next. Second, I introduce a method for learning hierarchical task networks (HTNs) from action sequences that helps AI2T induce simpler and more robust hierarchical programs than past systems. This approach is agnostic to action sequence lesson ordering, and induces HTNs with features like unordered groups and conditional actions that are useful for ITS rules.