Artistic expression

« various types of neural network.

Although classical and connectionist AI are often described as utterly distinctparadigms, research in both these approaches commenced because of this paper.

Early connectionist work wasfurther encouraged by McCulloch and Pitts in a paper of 1947.

They pointed out that the brain is a parallel-processing device, not a sequential one.

Moreover, it can function acceptably even when some cells misfire or die,or when the input signal is 'noisy'.

The perfect input data assumed within their first paper are, in real life, neithernecessary nor often available.

The question arises, then, how we (and animals) manage without them.

McCullochand Pitts described a statistical technique, based on differential equations like those of thermodynamics, whereby aparallel-processing system could compute (learn to distinguish) various patterns despite slight variations in theinput.

These (statistical) ideas were less biologically unrealistic than their earlier (logic-based) discussion.Nevertheless, the 1947 paper was less influential over the next three decades than their earlier work.

Only in the1980s did statistical, parallel-processing models achieve prominence (see §3 below). 2 Classical AI Classical AI is the best-known type of AI, and is sometimes called traditional AI.

It uses sequential programming (do this, then do that), and employs internal representations of lists, semantic networks, arrays andother information-processing structures.

These representational structures and their components are interpreted assymbolic representations of propositions and concepts (or beliefs and ideas).

Accordingly, this approach is alsocalled symbolic AI.

Most internal representations in classical AI are language-like, being constructed fromcomponents each of which has some distinct causal-semantic role (though just which role may vary according tocontext).

Some philosophers, such as Jerry Fodor, explain human mental states, or propositional attitudes, in termsof a hypothetical 'language of thought' having logical properties (compositionality, productivity, systematicity) likethose exploited in classical AI (see Fodor, J.A. ; Language of thought ).

A 'toy' example of one simple type of classical AI program (a production system) might look something like this: If thirsty then set goal to drink.

If current goal is drink and weather is cool then set goal to seek kettle.

If current goal is seek kettle and not in kitchen then go to kitchen and locate kettle.

If kettle is empty then fill kettle with water.

If kettle is full then put kettle on hob and heat hob and locate teapot.

(and so on) As this toy example suggests, every action, and every condition for action, has to be explicitly specified.

Actions that undo previous actions (such as emptying the kettle you justfilled) must be avoided.

Some unintended consequences of actions have to be anticipated and tidied up (turn offthe hob).

Default steps must be specified in case any precondition is not satisfied (hot weather, not thirsty).

Goal-subgoal structure must be recognized, and the program must be able to 'pop up' to the top goal-level when thelowest sub-goals have been achieved or abandoned.

Moreover (what the toy program does not show), proceduresmust be provided for carrying out the tests (is it cool, and is the kettle full?) and for executing the lowest-levelactions (going to the kitchen, locating and filling the kettle) (see Rationality, practical ).

Classical AI modelling is widespread in computational research.

It is used to study, for example, problem solving, planning, vision, robotics,learning, natural-language understanding, analogy and the perception and performance of music (Boden 1987 , 1988 , 1990 ; Rich and Knight 1991 ).

It is applied also to phenomena often assumed to be intractable for a computational (or even scientific) explanation, such as motivation, emotion and creativity.

Among the advantages of classical AIare its ability to represent hierarchical structure and to provide relatively transparent models (whose workings canbe well understood by inspecting the program).

A further advantage is that it can define 'strong' (exceptionless)problem constraints.

It is sometimes forgotten, especially by proponents of connectionism, that strong problemconstraints are often needed.

For instance, every sentence must have a noun phrase and a verb phrase; and waltztime in music demands that each bar have exactly three beats.

Admittedly, a composer may produce someanomalous bars (for example, having only two beats in the upper voice along with three in the lower); but onecannot keep doing this, or break out into march time, without abandoning the goal of composing a waltz.

Nor canone communicate intelligibly if one omits most noun phrases.

Given that certain rules are mandatory, an AI systemshould respect them, not approximate them by blurring them with others.

Although it began no earlier thanconnectionism, classical AI achieved visible success before parallel-processing models did.

The first major successesoccurred in the 1950s.

The logic theorist and general problem-solver of Newell and Simon introduced 'means-endanalysis', wherein a program analyses the problem as a hierarchy of goals and sub-goals (on indefinitely manylevels) and chooses the action most likely to reduce the difference between the current state and the desired state(the goal).

This method was widely adopted in theorem proving, problem solving and planning.

Another earlylandmark was Samuels' draughts (checkers) player, which played well and even learned to adapt to its opponent'sindividual style.

And early language-using programs used stored English word strings and simple linguistic schematato conduct 'conversations' in which the human interlocutors were occasionally (if briefly) persuaded that they wereinteracting with another person.

(See Feigenbaum and Feldman 1963 .) By the early 1970s there had been considerable advance.

For instance, natural-language processing could now be sensitive to highly complex syntacticstructure or to the unspoken assumptions hidden in the semantics underlying the actual words - so that programscould 'answer' questions about things not explicitly mentioned.

Machine learning was sometimes achieved throughthe program's having a model not only of the task domain but also of its own action strategies - which, withexperience, it modified.

Other advances followed.

Various high-level AI programming languages were developed,such as LISP ('list-processing language') and PROLOG ('programming in logic').

And Newell and Simon developed'production systems', a programming method based on if-then (condition-action) rules: if the condition is satisfied,then the action is taken.

The condition may be a complex conjunction or disjunction, including (sometimes) astatement of the system's current goal; similarly, the action may be complex and/or internal (see the toy AIprogram above).

These developments affected both technological and psychological AI.

Production systems, forinstance, are the core of most 'expert systems', but were originally proposed as a model of human thinking.

(Anexpert system is an AI program consisting of a set of 'If… then' rules, which can be used to aid human beings insolving specialist problems such as locating oil, planning a travel itinerary or diagnosing a disease.) 3 Classical AI and human thinking Traditional AI began with the assumption that symbolic logic is a normative. »