Thursday, August 5, 2010

A new approach to artificial intelligence


I'm intrigued by a 'bottom-up' approach to creating computer-based artificial intelligence. The Telegraph reports:

Computer-simulated life forms which reproduce themselves inside their electronic world have evolved to produce basic intelligence.

It is hoped that the discovery may in future lead to artificially intelligent brains "bred" within a computer.

The "Avidians", a race of digital beings in a computer world called Avida run by scientists at Michigan State University, with computer code instead of DNA that is copied - not quite perfectly - every time they breed. The random copying errors create differences in their code which dictate how well, or badly, they will perform in their simulated world.

Early experiments put the Avidians on a grid of cells, and let them live and die there. The grid had a gradient of food - cells at one end have more than the ones at the other, where the Avidians begin. After 100 generations of breeding, a mutation led to one of them evolving a "gene" instructing it to move forward. When it landed in a more food-rich cell, it reproduced more quickly, and had more offspring than its rivals.

After thousands more generations, the Avidians had evolved something more impressive: a rudimentary memory. They had started moving towards the food source in a zig-zag motion, changing direction when they were going in the wrong direction. To do that, they had to be able to compare their current cell to the previous one. Robert Pennock, one of the scientists behind the experiments, told New Scientist: "Doing this requires some rudimentary intelligence. You have to be able to assess your situation, realise you're not going in the right direction, reorient, and then reassess."

A later experiment added a new twist: cells that contained instructions on where to go to find food. Some of those instructions were simply "do what you did in the last cell". In order to make sense of those instructions, Avidians had to evolve a more complex memory - and duly did so. Laura Grabowski, another of the researchers, said: "The environment sets up selective pressures so organisms are forced to come up with some kind of memory use - which is in fact what they do."

This sheds some light on how intelligence originally evolved: MSU zoologist Fred Dyer says: "Laura's work suggests that the evolution of an ability to solve simple navigational problems depends on first evolving a simple short-term memory - and this in digital organisms that still don't exhibit something you would call learning." But the findings may, in the future, allow researchers to create true artificial intelligence.

Dr Grabowski says: "In the past, the approach has been to start with high-level intelligence and reproduce that in a computer.

"This is the opposite. We're showing how complex traits like memory can be built from the bottom up, from things that are really very simple."


There's more at the link.

This is fascinating! I worked with early artificial intelligence (AI) systems in the 1980's, using them to help with the design of commercial computer systems. The whole emphasis in those days was on 'top-down' development, taking the best approaches to a problem that human intelligence could produce and trying to synthesize their main features into a computer-based hierarchical decision-tree-based AI system. This approach turns the earlier procedure on its head. Instead of trying to synthesize human intelligence into a machine form, the system tries to have machine intelligence develop along the same lines that human intelligence must have used at the dawn of our species.

Of course, one of the problems with such a 'bottom-up' approach is that we don't know what sort of 'intelligence' will be produced. Sure, the scientists will probably attempt to 'steer' it in what they consider 'appropriate' directions . . . but what are the chances the Avidians (or their successors) might head off in a radically different direction, producing an 'intelligence' which is competent and effective by its own standards, but also completely alien to our way of thinking?

Or do we already have such an intelligence in our midst?




Hmm . . .



Peter

2 comments:

Crucis said...

I wonder just how many evolutionary iterations it will take to create a democrat?

That's one way to end the experiment---an evolutionary dead end.

LabRat said...

If one believes in a created universe, one might say this is simply making the entirely sensible move of modeling your approach off the Master Designer's. ;)

Less flippantly, I do believe such an intelligence would necessarily be quite alien; after all, it will have evolved for a completely different environment that operates by completely different rules and "senses" that perceive a completely different sort of reality.

Creating an artificial intelligence may actually be a much easier task than creating one that thinks like a human, which was what the first AI researchers were, IMO, trying to do.