Teaching AI Methods to Children

Linguistic and logical-mathematical intelligences, the first two intelligences in gardner83's definition, are crucial aspects of high-level thinking and intellectual processes. However, there are many other thinking abilities that are not explicitly represented. Therefore, we combine the first two intelligences, and give them a new name: intellectual intelligence, and add more components into it. Below is our definition of multiple intelligences.

1#1

The category of thinking strategies can be further divided into several subareas:

2#2

We believe that it is important to separate high-level thinking strategies from knowledge and language since they underlie other mental activities and operate upon domain knowledge. Such a detailed breakdown of thinking strategies can also help us to focus on specific and different thinking abilities, which directly correspond to various areas of AI. We can then teach, systematically, AI principles and methodologies in areas of AI to children to improve their thinking and problem-solving abilities.

Table 1 lists some of these thinking strategies and their corresponding AI principles and methodologies. A few examples of thinking puzzles, games, and educational software through which these thinking methodologies are taught are also included in the table.

 

Critical thinking Hypothesis-space reduction 20-question puzzles

(Ask key questions) Belief revision Lateral thinking puzzles Scientific investigation

Logical reasoning Inference rules (Modus Ponens, etc.) Minesweep, Mastermind

(Deduction) Proof by negation Three-muddy-boy puzzle

Rational reasoning and Probability estimation Betting, lottery

optimal decision making Utility theory Decision making in daily life

Inductive learning Occam's Razor Logic Journey of the Zoombinis

and analogy Complexity of concepts 1, 2, 4; what comes next?

Problem solving by search Search strategies, heuristics Sokowin, 8 puzzles

and planning Constrain satisfaction heuristics Maze, Cryptarithmetic

Problem-solving Divide-and-conquer Motor programming in Dr. Brain

strategies Recursion Hanoi Tower

As an example, lateral thinking puzzles are excellent for developing critical thinking skills and creativity. The instructor poses a puzzle (for example, a man is dead in a field with a unopened package next to him. Why did he die?), and students can ask only yes/no questions. The instructor answers these questions faithfully with yes, no, or irrelevant, according to the correct answer. The goal of students is to find out the instructor's solution by asking as few questions as possible. This fascinating game also resembles the scientific discovery and investigation process in which scientists design critical experiments to find out answers to some anomalies.

The AI principle for critical thinking is hypothesis-space reduction (as version-space reduction in machine learning): Every question should eliminate half of the hypothesis-space, no matter if the answer is yes or no. Children are then taught that it is not just half the number of the hypotheses, but half of the possibilities (some hypotheses may be much more likely than others). This teaches children about probability estimation and updating of likely events.

As we can see, these AI principles and methods are quite general since they are knowledge-independent, capable of solving various problems of the same kind. They also resemble meta-cognition (how to think) that educators often talk about and strive to teach to children. Acquiring such high-order thinking strategies that can be applied to other new problems in school and in life is the ultimate goal of learning.

In recent years, this educational and pedagogic paradigm of stressing the importance of learning how to learn instead of merely learning domain facts and rules of application has gained considerable supports in schools, and AI's role in education is also being expanded [Andriessen SandbergAndriessen Sandberg1999]. We believe that teaching AI for improving children's high-level thinking abilities is a fascinating new application of AI in education.

The AI principles and methodologies are also quite concrete, since they could be illustrated in detailed steps, just as computer algorithms, through thinking games and puzzles. Notice, though, that in most cases the correspondence is not at the algorithm level: it would be impossible to teach children (or adults) the exact A^* search algorithm, or any mechanical theorem-proving process. However, the principles can certainly be acquired and carried over to new tasks. For example, after learning principles of problem solving by search, a 9-year-old girl showed me that the 4-missionaries-and-cannibals problem has no solution by exhaustively examining the whole state-space (while avoiding repetitive states), even though the order of states searched is not as well organized as in the A^* algorithm.

I have been running a ``Creative Kids Workshop'' on weekends over the last two years for elementary-school children, and found that this fashion of top-down teaching of thinking skills very effective. Instead of just doing thinking puzzles (some schools do not even do that often) and hoping that children would develop high-level thinking skills themselves someday, children are directly taught thinking methods, which are illustrated repeatedly through different examples, puzzles, and games. It is my belief that this top-down teaching of thinking skills is much more effective than the bottom-up (problem-driven) approach.

It is also possible, although very difficult at present, to design AI tutoring systems that teach AI thinking methods. One main difficulty is the open-ended and interactive nature of the teaching environment (as the one in the Creative Kids Workshop). For example, when doing lateral thinking puzzles during the Workshop, unexpected questions are often asked, and clarification and disambiguation needed. This requires a large amount of common-sense knowledge. In addition, whether a question is good or not depends on the sequence of the questions. It is still very difficult to design an AI system that can answer such open-ended questions in unconstrained natural language.