Building Educational Technology Quickly and Robustly with an Interactively Teachable AI

  Interactive task learning (ITL) is a machine-learning paradigm that envisions AI that  can learn whole programs directly from the natural instructions of untrained users. 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 demonstra tions 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 learn 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 data efficient and robust interactive task learning, whereby correct and complete rule-based pro grams 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 esti mate 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.  

The machine learning and interactions design innovations of this work improve upon  the authoring-by-tutoring implemented in past systems like SimStudent and the Apprentice  Learner (AL) framework. I evaluate these improvements in two user studies each with 10  users. In study 2, half of our participants succeeded at teaching AI2T 100% complete and  accurate ITS behavior for two K-12 math domains when we evaluated the induced programs  on a large holdout set of 100 problems. These first-time participants worked with AI2T for  a median time of just 22 minutes per domain, half the time reported in our prior work.