LAUSR

Individualized tutoring has long been known as the most effective method of instruction. Bloom long ago suggested that there is a two-sigma improvement of learning with tutoring over other methods. However tutoring is uneconomical in most cases in terms of number and time: there are simply not enough qualified tutors available; even school teachers who would like to embrace individualized instructions find that the huge time burden makes it impractical. A tantalizing answer to this shortage is computer tutors. During the first heyday of Artificial Intelligence in the 1980s, Intelligent Tutoring systems were proposed to engage in Socratic tutoring. Hopes were raised when qualitative models for physics were created, model tracing programs developed, and misconceptions in many fields were analyzed. Within well defined areas, like algebra, significant progress was made. However, it was soon obvious that students not only had organized misconceptions but also there were many undefined errors, and it was not simple to distinguish among them. The first order predicate logic systems could not deal with human inconsistencies. Rule based AI was inadequate, and statistical techniques were radically inadequate. There are two key conditions which must be met in realizing effective computer tutors: a model of the curriculum and it is all about knowing each individual learner. One technology requirement is how to keep a timely and complete record of the learning process and another one is how to analyze and identify learning problems/obstacles for each individual learner. The development of smart classrooms and online learning platforms help the first condition. For the second condition, in the forty years since, new statistical approaches to AI, using many machine learning techniques such as Support Vector Machines and latent semantic analysis have matured rapidly, and have reached a new critical mass with deep learning, neural networks, and convolution that afford opportunities for new learning analytics. The papers collected in this field spread across many disciplines and statistical approaches. The vibrant research that these papers highlight suggests a new beginning for Intelligent tutoring systems that may still take years to unfold but provides a bright hope for what once seemed nearly impossible. One paper examines a rule based approach to understanding how students game an ITS with augmentation frommachine learning technologies. Another examines the common errors made in thousands of Block Based Programming lines of code. One of the biggest areas for using big data and learning analytics is in MOOCS where thousands of students take the same online course. Standard factor analyses still work for these, but big data analyses may make large improvements. Hidden Markov models can reveal interaction patterns for low and high achieving students. Online forums contain much text that needs to be analyzed to understand students’ issues and answers. Machine learning models of creativity, emotion and knowledge can be very helpful. Utilizing automatic text analysis, another study built a hierarchical linear model that examines the influence of the pacing condition of a massive open online course (MOOC), whether it is self-paced or instructor-paced, on the demonstration of cognitive processing in a HarvardX MOOC. Text analysis can also be used to confirm states of confusion in students in MOOCs.