What is AI (2)?

The only true definition is ``that which AI people do''. Looking at the early history of AI, there were already a number of different approaches by the end of the 1950s. For some examples and commentary, see Chapter 8 of The Guide to Expert Systems, by Alex Goodall (Learned Information, 1985; RSL Comp BD 36; PSY KH:G 061). Note:

• The cybernetic approach. Derived from cybernetics, the study of control and self-organising systems. The idea was to put together a very large number of simple and similar elements, and study the way in which (hopefully) the system ``learnt'' to respond to perturbations in its environment by acting so as to negate their effect. Intelligent behaviour was a by-product of this internal goal-seeking or homeostasis. See Ross Ashby's homeostat.
• The neural-net approach. The elements are simulated neurons. Then usually simulated in hardware; now, in software. Many different levels of biological fidelity are possible. See Grey Walter's turtles, and the Perceptron.
• The symbolic or programming approach. One writes a symbol-manipulating program for a digital computer. Exemplified by the Advice Taker.

So right at the start, there were different approaches and methods. This has persisted. AI has few unifying principles and lots of disagreements about its objectives and methods. Too many practitioners spend the time mud-slinging, rather than trying to understand different approaches and see what they have in common.

Like psychology, AI is too large for any one person to claim that their work spans it all. Most researchers work on well-defined subtopics, and some do not see themselves as having much connection with the rest of the subject. For example, Steve Muggleton in the Programming Research Group works on ``Inductive Logic Programming'' - a form of machine learning where a computer program tries to learn general logical rules from specific instances. Although any complete artificial intelligence would need to be able to learn, Steve never claims to be working on AI. Like ILP, neural networks and genetic algorithms are also kinds of machine learning, but it would be rare to find an individual whose work covered more than one of these three topics. In fact, there are many different designs of neural net and genetic algorithm, and most researchers will specialise in just one kind.