Constraints and Optimality

Sometimes early neurological or morphological constraints can actually have positive effects in the long term; for example, in the case of Elman's neural network model of language learning, initial limitations on active maintenance abilties actually resulted in better corpus learning relative to networks that were not initially constrained. It appears that the constraints on motor skills in human infants may actually be a very important part of our eventual mastery at controlling our own bodies.

Nearly half a century ago, Bernstein realized that human infants learn many motor movements (such as walking, or grasping) first by locking their joints, so as to constraint the number of degrees of freedom they have to work with. After practice, the degrees of freedom are iteratively unfrozen beginning with those joints farthest from the infants' center of gravity.

Lungarella and Berthouze modeled this developmental progression in a several autonomous robots, one of which is pictured at the start of this article. They were able to show that motor performance is better in those robots that are given progressive control over their degrees of freedom than in those robots which are allowed to use all available degrees of freedom from the start. This same pattern was observed for a variety of motor skills, including the pictured simulation of a "jolly jumper" in which infants learn to bounce. Developmental constraints, such as the number of available degrees of freedom, may be an important mechanism in the eventual emergence of optimal behaviors.

The developmental approach to robotics is relatively new and unexplored. For example, although there are developmental robotic implementations of basic motor skill learning (reaching, grasping, gazing) there are none of more complex motor skills, such as running, walking, or dancing. Those systems that are capable of bipedal motion have been explicitly designed for bipedal motion, as opposed to being designed to learn.

Given the difficulty in traditional techniques for programming bipedal motion, one wonders whether developmental approaches might be better for other "hard AI" problems as well. To quote Lungarella et al.: "the designer should not try to engineer ‘intelligence’ into the artificial system (in general an extremely hard problem); instead, he or she should try to endow the system with an appropriate set of basic mechanisms for the system to develop, learn and behave in a way that appears intelligent to an external observer."


Blogger Mentifex said...

In gentle contradiction of Lungarella et al. (op. cit.), the time has come to think and discuss ways to engineer ‘intelligence’ into the artificial system -- and specifically into the constraints that will engender optimality in the motor system and other mind modules.

3/09/2006 08:28:00 AM  
Blogger Chris Chatham said...

Hi Arthur - I had written a reply to this, which seems to have disappeared.

First, I wanted to say that we appear to agree on the idea that constraints can be important for learning. I'm not sure if we disagree on anything, actually, but just in case, I'll hammer my view home ;)

I would emphasize that the important thing to do is not to engineer intelligence, but to engineer the capacity for learning and growth. After all, intelligence is created every day from fairly stupid entities - aka babies - so why not use the "system principles" guiding child development to guide our own attempts at developing intelligence?

I agree that motor skills is a particularly good demonstration of the failure of traditional AI approaches. Certainly, it's not an easy task for children either: think about how long they spend on their back, simply wiggling their toes, before they even begin to crawl! Then, how long they crawl before they begin to stand up. And finally, how long they wobble before they're fully capable of balanced walking. All the while, the motor cortex is developing.

I believe these developments are slow (and synced) for a reason; the contrained degrees of freedom, and the contrained cognitive architecture, work hand-in-hand to allow fully capable behavior to eventually emerge.

I think the approach of trying to design an AI system to be "smart from the start" is misguided. Instead, we should be building biological simulations, with an engineered ability to learn and grow - through self-induced changes in architecture, input, and behavior.

This developmental approach has three advantages: one, we know that it works (biology does it somehow!) Second, we have a working prototype version (ourselves!) And finally, if we are successful in developing such a system, it's likely to be much more flexible than a system that was designed to be "smart from the start."

3/09/2006 02:40:00 PM  
Blogger Mentifex said...

When you talk about babies and "how long they spend on their back, simply wiggling their toes, before they even begin to crawl," you bring to mind my attempt at theorizing that the babies are experiencing random firings of their motor memory neurons, with the result that an ever widening feedback loop allows the babies to take conscious, volitional control of their motor behavior.

In the rest of your comment-in-response you endorse bottom-up mind-design strategies in preference to the top-down AI design that I have been working on. As Nature has shown, your way is guaranteed to work, but for me it takes too long. Luckily,
since people work from both points of view, some pretty interesting results are beginning to "emerge" (another "loaded" word :-).

3/10/2006 05:13:00 AM  
Blogger Chris Chatham said...

I like your theory about the babies exploring the "motor space" via essentially random firings.

And I think you're exactly right on the point about people working from both points of view. I think that ACT-R may be a good example of that.

Good luck with your sourceforge project - I am unfortunately not well versed enough in either the brain or computer programming to be of much help.

3/10/2006 07:52:00 AM  

Post a Comment

<< Home