Current Topics in Artificial Intelligence

Artificial Intelligence (AI) technology has been extended for use in tutoring systems to dynamically customize material for individual students. These techniques model and reason about the student, the domain and teaching strategies, and communicate with the student in real time. Evaluation results show increased learning, reduced costs, and improved grades. We will demonstrate intelligent and distributed technology that makes education available anytime and anyplace. At the grade school level, a mathematics tutor positively influences students’ confidence and image of their mathematics ability. Machine learning was used to model student performance and to derive a teaching policy to meet a desired educational goal. At the college level, an inquiry tutor moves students towards more active and problem-based learning. We will also discuss other tutors that introduce new pedagogy and address inequities in the classroom. 1 Artificial Intelligence in Education The field of Artificial Intelligence in Education (AIED) is relatively new, being less than thirty years. Broadly defined, AIED addresses issues of knowledge and learning and is not limited solely to production of functional intelligent tutors. Issues and questions addressed by this field include: What is the nature of knowledge? How is knowledge represented? How can an individual student be helped to learn? What styles of teaching interactions are effective and when? What misconceptions do learners have? The field has developed answers to some of these questions, and artificial intelligence (AI) techniques have enabled intelligent tutors to adapt both content and navigation of material to a student’s learning needs. The goal of AI in Education is not to reproduce existing classroom teaching methods. In fact, in some cases the goal is to remove the traditional education mold altogether. As working and learning become increasingly the same activity, the demand for lifelong learning creates a demand for education that will exceed the capability of traditional institutions and methods. This creates an opportunity for new intermediaries and learning agents that are not part of the traditional, formal education system. Such opportunities are likely to be supported by computer technology. 2 Beverly Park Woolf Fig. 2. Interface of AnimalWatch for a simple addition of whole numbers problem. Fig. 1. Real World Context: In AnimalWatch, the student chooses an endangered species from among the Right Whale, Giant Panda and Takhi Wild Horse. Ample evidence exists that intelligent tutors produce a substantial improvement in learning and productivity in industry and the military. Formal evaluations show that intelligent tutors produce the same improvements as one-on-one human tutoring, which increases performance to around the 98 percentile in a standard classroom [1]. These tutors effectively reduce by one-third to one-half the time required for learning [2], increase effectiveness by 30% as compared to traditional instruction [3, 2, 4], and networked versions reduce the need for training support personnel by about 70% and operating costs by about 92%. The term “intelligent tutor” designates technology-based instruction that contains one or more of the following features: generativity, student modeling, expert modeling, mixed initiative, interactive learning, instructional modeling and self-improving. The key feature is generativity – the system’s ability to generate customized problems, hints or help – as opposed to the presentation of prepared “canned” instruction. Generativity relies on models of the subject matter, the student and tutoring, which enable the tutor to generate customized instruction as needed by an individual student. Advanced instructional features, such as mixed-initiative (a tutor that both initiates interactions and responds usefully to student actions) and self-improving (a tutor that evaluates and improves its performance as a result of experience), set tutors apart from earlier computer-aided instructional systems. No agreement exists on which features are absolutely necessary and it is more accurate to think of teaching systems as lying along a continuum that runs from simple frame-oriented systems to very sophisticated intelligent tutoring. The most sophisticated systems include, to varying degrees, the features listed above. For example, the Arithmetic Tutor described below was generative since all math problems, hints and help were generated on the fly based on student learning needs (Figs. 1-3). The tutor modmodeled expert knowledge of arithmetic as a topic network, with nodes such as “subtract fractions” or “multiply whole numbers” which were resolved into child nodes such as “find least common denominator” and “subtract numerators,” Fig. 4. The tutor Reasoning about Teaching and Learning 3 Fig. 3. A sample pre-fraction problem. modeled student knowledge, recording each sub-task learned or needed based on student action and the tutor was self-improving in that it used machine-learning techniques to predict a student’s ability to correctly solve a problem. 2 Customizing Help by Gender and Cognitive Development The first example tutor, AnimalWatch, used AI techniques to adapt its tutoring of basic arithmetic and fractions, Figs. 1-3. It helped students learn fractions and whole numbers at a 4-6 grade level. The tutor used student characteristics including gender and cognitive development, and an overlay student model which made inferences about the student’s knowledge as he/she solved problems. The tutor adjusted its problem selection to provide appropriate problems and hints. For example, students unable to handle abstract thinking (according to a Piagetian pre-evaluation) benefited from concrete representations and concrete objects to manipulate instead of formal approaches, equations, or symbols and textual explanations, Figs. 5-6. Students moved through the curriculum only if their performance for each topic was acceptable. Thus problems generated by the tutor were an indication of the student’s mathematics proficiency and the tutor’s efficiency as described below. Results indicated that girls were more sensitive to the amounts of help than to the level of abstraction (e.g., the use of concrete objects to manipulate, Fig. 7, vs. equations and procedures, Fig. 6) and performed better in problems when the help was highly interactive. Boys, affected by the level of abstraction, were more prone to ignore help and to improve more when help had low levels of interactivity. AnimalWatch tutored arithmetic using word problems about endangered species, thus integrating mathematics, narrative and biology. Math problems were designed to motivate students to use mathematics in the context of practical problem solving, embedded in an engaging narrative, Figs. 1 and 2. Students “worked” with scientists as they explored environmental issues around saving endangered animals. AnimalWatch maintained a student model and made inferences about the student’s knowledge as s/he solved problems. It increased the difficulty of the problems depending on the student’s progress and provided mathematics instruction for each student based on a dynamically updated probabilistic student model. Problems were dynamically generated based on inferences about the student’s knowledge, progressing from simple one-digit whole-number addition problems to complex problems that involve fractions with different denominators. 4 Beverly Park Woolf Fig. 4. A Portion of the AnimalWatch topic network. The student's cognitive level was determined via an on-line pretest [5] based on Piaget's theory of development and a series of questions on topics such as combinatorics, proportions, conservation of volume and other elements that determine student ability to reason abstractly. Several evaluation studies with 10 and 11year-old students totaling 313 children indicate that AnimalWatch provided effective individualized math instruction and had a positive impact on students' own mathematics self concept and belief in the value of learning mathematics [6, 7]. When a student encountered a difficult problem, AnimalWatch provided hints classified along two dimensions, symbolism and interactivity. The first hints provided little information, but if the student kept entering wrong answers, AnimalWatch provided hints that ultimately guided the student through the whole problem-solving process. Expert Model. The expert model was arranged as a topic network where nodes represented skills to be taught, Fig. 4. The links between nodes frequently represented a prerequisite relationship. For instance, the ability to add is a prerequisite to learning how to multiply. Topics were major components of the curriculum, e.g., “add fractions” or “divide wholes”, while skills referred to any curriculum elements (including topics), e.g., “recognize numeration” or “recognize denominator.” Subskills were steps within a topic that the student performs in order to accomplish a task. For example, the topic “adding fractions” had the subskills of finding a least common denominator (LCM), converting the fractions to an equivalent form with a new numerator, adding the numerators, simplifying the result, and making the result proper. Table 1. Three sample add-fraction problems and the subskills required for each.