Comenzaba yo a desesperar con el número de noviembre del 86 de la revista Byte: no ha sido hasta la página 225 que he encontrado algo reseñable…
![MACHINE LEARNING
BY Angelos T. Kolokouris
One approach to teaching computers to learn is with a language, such as Marvin's, that grows
IN AN ATTEMPT to make computers more accessible to humans, many researchers in the artificial intelligence field have been exploring ways to enable computers to learn. The motivation for machine learning is to have computers extract concepts and relations from databases or through interactive sessions with a user and then use them in any knowledge-intensive activity. Developing knowledge bases for expert systems applications is one such activity. Studying computer-based learning techniques will also give us a better understanding of our own mental processes.
A common method by which a machine learns is by proceeding from specific instances to general rules that more economically capture the content of the given instances. This form of inductive generalization is characteristic of learning from examples and improves on the performance of a knowledge-processing system by condensing a large base of its knowledge into a considerably smaller one. This reduction in size results in a more efficient search of the knowledge base.
When you take a closer look at the programs that learn from examples, you can distinguish two types: those,
called data-driven learners, that generalize by relying entirely on the data presented to them, and a group of more elaborate programs, called model-driven learners, that proceed by generating fairly general hypotheses that are subsequently tested against the given examples or against the user in a typical interactive session. In what follows I will contrast the model-driven learner with the data-driven learner and give an example of the former using a model-driven learner called Marvin. | Editor's note: The Prolog source code for Marvin is available on disk, in print, and on BIX; see the insert card following page 352 for details. It is also available on BYTEnet; see page 4.]
What Is Learning?
To give you an idea of just what's involved in modeling learning with computers, I should examine just what learning is. The best definition of learning that I am aware of— that of Herbert Simon of Carnegie-Mellon University (reference 1)— is that 'learning denotes changes in the system that are adaptive in the sense that they enable the system to do the same task or tasks drawn from the same population more efficiently the next time." 'Inking a closer look at this definition, however, you may realize that it deals predominantly with skill acquisition or improvement, but not all learning is so concerned. Learning systems are those that are able to extract knowledge from raw data or through intersystem informative exchanges, including conversations with the user. Learning systems ought to have the ability not only to acquire knowledge in a cumulative form but also to absorb it.
Machine Learning Strategies
The methodologies used in machine learning applications with some degree of success are rote learning, learning by being told, learning by analogy, and learning from examples. In rote learning the computer makes no special effort to learn. The elements of new knowledge are given to it, through programming or by accessing an external data file, and inferencing...](https://i0.wp.com/obm.corcoles.net/wp-content/uploads/2026/10/image-1.png?resize=723%2C1024&ssl=1)
…pero ahí lo tenéis, un «no estamos haciendo nada nuevo» de manual, pero cuarenta años de ley de Moore dan para diez mil veces más potencia…
(El Marvin al que hace referencia el artículo, de hecho, ya era «mayorcito» para entonces: aparece en la literatura científica en 1981 (PDF) , en lo que diría que es una tesis doctoral. Y el concepto de machine learning, de hecho, es de 1959.)
Como decía, el número de noviembre del 86 de Byte no es una maravilla, pero si os apetece darle un vistazo, tenéis una review del Mac Plus, con sus diskettes de doble cara y 800 kilobytes y un descomunal mega de RAM (ocho veces más que el Mac original). Había, además, kits para añadir un mega más, a partir de 295 dólares de la época (algo más de seiscientos, hoy, con la inflación). No haré comentarios sobre los datacentres, el RAMageddon y el precio de la RAM, pero ni así (¡espero!) vamos a volver a los costes de hace cuatro décadas.
Y no me voy sin apuntar a esta review del SPSS/PC+, aunque solo sea para destacar que el Statistical Package for the Social Sciences debe ser la aplicación de software más longeva de la historia (como mínimo, de las que yo conozco, lo es de largo): dice la Wikipedia que se lanzó en… 1968, en la época de las tarjetas perforadas.
El mes que viene, más (y más interesante, espero). O no.