dandan | 12 Agosto, 2008 23:19 |

Estos días he estado leyendo sobre dos investigaciones que se han hecho a partir de procesos emergentes, es decir, conseguir resultados homologables con la realidad de llegada a que estamos acostumbrados sin colocar directamente ese destino en el diseño de salida, sin forzar el proceso en esa dirección. Intentar que llegen a destino por un proceso de autoorganización de abajo arriba: no colocar primero el circuito en el que tienen que correr los coches, sino soltarlos todos por el campo y que acaben corriendo en el circuito. Los dos experimentos se han hecho en ámbitos distintos, pero el proceso es muy similar. A ver, a ver.
La primera trataba de estudiar la creación de un lenguaje artificial (una demo de prueba) a partir de un conjunto de señales sin ningún sentido inicial. El interés en este caso consiste en que este tipo de procesos emergentes se han estudiado fundamentalmente en biología (autoorganización celular en organismos o de organismos-hormiga en hormigueros, por ejemplo), pero si se repetía el mismo proceso con el lenguaje, entonces se encontraba el mismo patrón en una teoría de la evolución no-biológica, sino ya cultural y por tanto plenamente humana. Y esto tiene tela, porque existe la idea muy extendida de que una cultura no se desarrolla sin señales de circulación (stop, ceda el paso...), pero bueno, poco a poco. El experimento lo explican así:
Kirby and his team showed people a collection of pictures paired with gibberish words, and later tested which pairs they could recall. Whether or not the recollections were accurate, they were recorded and used as the basis of the next group's language training. As the process was repeated, patterns emerged: a certain word might be used, for example, to describe anything that moved horizontally, and another to indicate objects that bounced.
The language that emerged from the first set of iterations, said Kirby, was limited and simplistic. But for the next set, they discarded duplicate words. Confronted with this selection pressure -- analogous, perhaps, to that exerted by nature on hunters with few words for their prey -- the language became precise and highly structured.
Structure, said Kirby, was the key to a language being remembered.
"Over many generations, the grammar goes from ad-hoc and inexpressive into a language that's cleanly structured and expressive," he said. "But what's evolving here isn't the agents" -- the speakers -- "but the language itself. It has its own evolutionary imperative. It wants to be passed on, and finds ways of doing that. We're its hosts."
No acabo de ver muy claras esas conclusiones (me suenan bastante a Chomsky): descubrir "the language itself" al final de un proceso emergente desarrollado por agentes suena un poco raro, pero bueno, cada cual que llame como quiera a lo que se transmite, mientras acepte que se puede seguir modificando por otros agentes posteriores (y esto no acaba de ser "itself"). Me parece más interesante la carga final:
He continued, "But people who are interested in culture more generally might take this work and study the emergence of design in a lab. I'd like to see how far that can be pushed. What kinds of adaptations would a culturally evolving practice come up with? How much of what happens around us, that appears rational and intelligently designed, is the product of a blind process?"
Vale, pues la segunda investigación no tiene que ver con el lenguage, sino con la robótica. Conseguir que un robot se mueva como nos movemos nosotros, que se detenga ante una pared, que suba escaleras o que baile con la música, les ha dado mucho trabajo a los ingenieros últimamente. Sabemos como lo hacemos nosotros (+ o -) pero no está tan claro como reproducirlo. Pues el experimento ha consistido en olvidar como lo hacemos nosotros (línea de llegada) e intentar conseguir que unos robots virtuales lo aprendan por ellos mismos igual que lo aprende un niño. Se trata de una simulación en pantalla, pero los resultados son espectaculares. Lo explican así:
Prof Ralf Der at the Max Planck Institute for Mathematics in the Sciences in Leipzig and the team at Edinburgh's School of Infomatics has applied the software to simulated animals and humans that are not given any explicit instructions on how to move. Nor do they know anything about the virtual environment.
All the neural network brain of each creature has to work with when controlling the humanoid is the angle of the 15 joints of the virtual creature. Dogs and snakes have up to 25 so called degrees of freedom.
The clever part is that the network is controlled by a process called "self regulation".
A deep mathematical analysis of living things has come up with an insight into why they are special, dubbed the "edge of chaos" by American computer scientist Chris Langton, to sum up how living things have the right balance between blindly following rules and creatively adapting to new situations.
The self-regulation principle drives the behaving robot into behaving in this special manner in two ways. "On the one hand actions should be the most sensitive answers to the current sensor values," said Prof Der.
This gives rise to a kind of "butterfly effect", he said, referring to how the effect captures the essence of chaos: a butterfly flapping its wings in London can, in principle, cause a subsequent hurricane in the Philippines because the system - in this case the Earth's atmosphere - is so "sensitive" that if there is even the slightest uncertainty in measuring the current weather conditions, then the weather in a few weeks' time is unpredictable.
For Prof Der's work, a tiny perturbation in measurements made by the sensors leads to large variations in the actions of the robot. But the robot is still under the control of physics and subject to the additional requirement that behaviour has to be predictable.
As a consequence of trying to meet both objectives the virtual creatures produce spontaneous behaviours "which are its most natural ways to move and act," he said.
While being active the robot is learning about itself and its surroundings at the same time.
| « | Agosto 2008 | » | ||||
|---|---|---|---|---|---|---|
| Lu | Ma | Mi | Ju | Vi | Sa | Do |
| 1 | 2 | 3 | ||||
| 4 | 5 | 6 | 7 | 8 | 9 | 10 |
| 11 | 12 | 13 | 14 | 15 | 16 | 17 |
| 18 | 19 | 20 | 21 | 22 | 23 | 24 |
| 25 | 26 | 27 | 28 | 29 | 30 | 31 |