Dawson Margin Notes On

Pfeifer and Scheier, Understanding Intelligence

Chapter 2: Foundations of Classical Artificial Intelligence and Cognitive Science

The purpose of this chapter is to review the classical approach to AI. “With an understanding of classical AI, the reasons for the embodied cognitive science approach are much easier to see.”

2.1 Cognitive Science: Preliminaries

Two key historical meetings launched classical AI. The first was the 1956 symposium on information theory, and the second was the 1956 Dartmouth conference. “The discussions at both conferences centered on what came to be called artificial intelligence and information processing psychology: the analogy between human thinking and processes taking place in a computer.” One consequence of this view was the sense-think-act cycle, with the emphasis on the internal “think” component. “In sum, the core idea that emerged from all this work was the belief that complex processes are required for transforming the stimulus (input) into the response (output).

Cognitive science is, by definition, interdisciplinary. How is communication possible? “In traditional cognitive science and AI, the language of information processing kept the field together. It is in fact more than a mere language. It endorses basic beliefs about the nature of intelligence. ‘Computation’ and ‘representation’ are the key words that best characterize these beliefs. NB: If these beliefs are so powerful to be strong unifying forces, then we want to be sure that there are lots of advantages offered by any alternative proposed to replace them!

2.2 The Cognitivistic Paradigm

What are the basic concepts of cognitivism? First, cognition is computation, as formulated in an account of the Turing machine. Turing machines are universal information processors, and are therefore psychologically interesting (Church-Turing thesis). One problem with this view, raised to foreshadow the next chapter – does the physical realization of a UTM lead to engineering problems (e.g., tape handling) – scaling up problems.

Second basic concept is functionalism – “intelligence or cognition can be studied at the level of algorithms or computational processes without having to consider the underlying structure of the device on which the algorithm is performed.” That is, emphasis on software instead of hardware, as in the proposal of the physical symbol system, and the view that intelligence involves symbol manipulation. In this view, processes act upon representations or symbol structures, and these representations reflect the real world.

Functionalism leads directly into the possibility of using computers to simulate cognitive processing. Pfeifer and Scheier have no problem with this. However, they are critical of “the analogy between human thinking and processes running in a computer, that is, information processing as the manipulation of symbols.” NB: Again, there are lots of interesting successes of this type. Can Pfeifer and Scheier provide us with examples that will compete?

2.3 An Architecture for an Intelligent Agent

To illustrate some of the basic ideas of classical cognitive science, Pfeifer and Scheier consider a design issue – what kinds of abilities would have to be built into an intelligent agent? They provide the following ideas:

• Vision, for instance in Marr’s sense – the reconstruction of a 3D representation of the world on the basis of light information
• Learning – the transfer of previous experience to novel situations, as defined in classical systems such as decision trees
• Memory – the proposal for different memory stores, and considering the symbolic formats in each of these
• Control – determining which operation to perform next, usually governed by some goal-directed principle such as means-end analysis, which can lead to problems with a combinatorial explosion (NB: problem foreshadowing again!)
• Other tasks, such as communication are likely to be required too

Considering all of these, Pfeifer and Scheier characterize the classical approach in terms of the following 7 design principles

1. Model as computer program
2. Goal-based designs
3. Rational agents
4. Modularity
5. Sense-think-act cycle
6. Central information processing architecture
7. Top-down design

NB: Are these the design principles followed by the classical approach? What are there advantages?