Knowledge-based systems

Artificial intelligence (AI), with its ability to tackle different classes of problems from those dealt with by conventional computing, has seen its economic importance grow considerably. These problems relate to human activities commonly associated with intelligence (perception, decision-making, planning, diagnosis, data interpretation, language comprehension, design), and have the common feature of requiring the reasoned exploitation of a vast amount of knowledge, most of it specific to the field under study and acquired from experts.

At the same time, AI has entered popular culture, notably through numerous fiction books and films, and games often set in virtual worlds. A number of recent successes, such as the Deep Blue chess program's victory over world champion G. Kasparov and the Mars Rover mission, have also raised the profile of certain aspects of AI.

As soon as the computer appeared, Alan Turing and other researchers hypothesized that it was possible to automate reasoning using general algorithms based on a set of logical rules applied to symbolic structures, following on from work on mathematical logic. Formal systems have shown their intrinsic limits (particularly with the work of Gödel and Church) for modeling reasoning. The need to restrict reasoning to a well-defined field of application, and to support this reasoning with knowledge of various kinds, thus emerged very early on in AI. This symbolic approach to AI gave rise to knowledge-based systems.

Another approach, which can be described as connectionist, emerged concomitantly with the beginnings of AI in the 1950s. It is based on the functioning of the cerebral cortex. The basic entity is a model of the neuron, a system being formed by the interconnection of a large number of such "neurons" (the most commonly used model is the formal neuron proposed by McCulloch and Pitts in 1943). This is a very rudimentary model of neuron function, in which the accumulation of the neuron's synaptic activities is ensured by simple weighted summation. The interconnection of a set of such units provides a neuromimetic connectionist system, also known as a neural network, with a number of interesting properties. Chief among these is the ability of the network to learn from examples. Such networks are used in a variety of fields: optical reading of texts, postal codes or signatures, fault diagnosis, quality control, industrial process control, stock market estimates, weather forecasting, etc.

In addition, probabilistic and statistical models are increasingly being used to account for the variability of the phenomena under study. This is particularly the case in automatic pattern recognition (speech, written characters, etc.), but also in reasoning, notably with Markov...