Artificial Intelligence in Education

The field of artificial intelligence and education is grounded in three academic disci-plines: computer science, psychology, and education, which all contribute to the de-velopment of the interdisciplinary field of Intelligent Tutoring Systems (ITS) [Woo08].

Research on intelligent tutoring serves two goals. Beside the obvious goal of de-veloping systems for automating education, an equally important goal is to explore epistemological issues concerning the nature of the knowledge that is being tutored and how that knowledge can be learned [ABCL90].

The idea of using computers to enhance learning began with the Computer-Aided (or Assisted) Instruction (CAI) systems (e.g., see [KBW83]), where computers were used mostly for presentation of student material. The main deficiency attributed to these systems is their static behavior; they are unable to interact with students or adjust to the specific student needs. Simple computer assisted instruction systems 14

1.5. Related Work

suffer from the fact that in general they do not know the subject matter they are teaching. Intelligent tutoring systems use artificial intelligence (AI) formalisms to represent knowledge in order to improve on CAI systems [Yaz86].

One-to-one tutoring with personal human tutors provide a highly efficient learn-ing environment and have been estimated to increase mean achievement outcomes by as much as two standard deviations [Blo84]. Education based on CAI systems has also been well documented to improve learning at the elementary, secondary, higher-, and adult-education levels. A meta-analysis of several hundred well-controlled stud-ies showed that student scores increased by 10% to 20%, the time to achieve goals decreased by one-third, and class performance improved by about one-half standard deviation. The current state-of-the-art intelligent tutoring systems are estimated to increase mean achievement outcomes by about one standard deviation [Woo08].

To our knowledge, intelligent tutoring systems that have been most successful at aiding student learning are Model-Tracing Tutors (MTT) [AP91] that are com-monly used in teaching problem solving domains and allow the tutor to follow the problem-solving steps of the student through the use of a detailed cognitive model of the domain. MTTs have had considerable success in improving student learning [ACKP95]. Carnegie Learning, a company founded by researchers from Carnegie Mellon, produced the commercial version of such tutor for use in high school math-ematics classes, which was used in about 10% of the U.S. high school math classes in 2007 [Woo08]. A second successful and widely used model-tracing tutor is the Andes Physics Tutor [VLS⁺05].

The core of model-tracing tutoring systems is an expert module that contains the cognitive model of the domain. Such cognitive models are usually based on a theory of human cognition, for example, the Carnegie Learning tutors are based on ACT-R, a learning theory and cognitive architecture framework [And93b]. ACT-R assumes that skill knowledge is initially encoded in a declarative form when students read or listen to a lecture. Students employ general problem-solving rules to apply declarative knowledge (concepts, facts, procedures etc.), but with practice, domain-specific procedural knowledge is formed. ACT-R assumes that procedural knowledge can be represented as a set of independent production rules that associate problem states and problem-solving goals with actions and consequent state changes.

Several research issues limit the use of Model-Tracing Tutors. Production rules have limited generality, and all model-tracing tutors suffer from the difficulty of ac-quiring problem-solving models, which requires cognitive task analysis, an enormous

1. INTRODUCTION

undertaking for any nontrivial domain. Cognitive analysis is typically performed manually and is tedious, time consuming, and error prone. Student models are of-ten hand-coded and remain fossilized unless exof-tended with human help. Addition-ally, this method is nearly impossible to reproduce for disciplines in which no well-developed psychological theory exists, such as medical diagnosis or law [Woo08].

Thus, whereas intelligent tutoring are proving to be useful they are also difficult and expensive to build [Mur99], mainly because building the expert model is difficult.

The MTT implicitly require complete domain knowledge, which requires a lot of knowledge engineering. And although many authoring tools (tools for building an ITS without the need of a programmer) were proposed, they have not been shown to be usable for modeling domain expertise [Mur99].

Building the expert module of a tutoring systems is similar to building the knowl-edge base of an expert system, where machine learning is commonly used as an alter-native way of obtaining the expert knowledge [FR86]. However, it should be stressed that the kind of knowledge required for an ITS (including the domain knowledge component) is different to that required for an expert system in the domain [Cla87].

It is quite possible to have an expert system that can perform the task well but that is poor at teaching or even explaining its reasoning because so much knowledge re-mains implicit [Twi92]. While it was shown that machine learning can be successful in building knowledge bases for expert systems [LS95] in terms of performance, the major problem with this approach is that these models usually do not mimic the cog-nitive processes (how an expert or a student solves problems in the given domain), which is the most important requirement of the expert module. Sison and Shimura [SS98] shed light on the difficulties of using machine learning for automating the construction of student models as well as of the background knowledge necessary for student modeling.

In the construction of intelligent tutoring systems, the acquisition of background knowledge, either for the specification of the teaching strategy, or for the construction of the student model, identifying the deviations of students’ behavior, remains one of the unsolved problems [Ant08]. A still unanswered question is: is it possible to conceptualize (semi)automatically the domain in a way that conforms to the way the experts want to have their knowledge organized and presented?

16