We can see in fact hints at the need for externalized sensory capability in an intelligence when we look at the cutting edge techniques in robotics. In robotics the problem was approached for nearly 40 years with the idea that you can computationally determine all the necessary motions of external limbs in order to dynamically balance and ambulate walking robots...but it turned out that doing that top down approach was not only extremely difficult, requiring massive amounts of computational power...but it was also doomed to always be less efficient from a power utilization stand point as well as simply not nearly as dexterous for free flowing motions (once where a precise path of motion is not followed).
The pieces started coming together in the mid 80's and came out of the work of people at MIT who started thinking seriously about asking how nature does these things. The Leg lab was famous for producing attempts at ambulation that mimicked animals and insects in various ways to reduce the necessary degrees of freedom and thus enable reduced computation for ambulation with steady balance but these were still not getting it right.
The key insight came from Marc Raibert then at the Leg Lab, he reasoned quite obviously that if it is true that ants and roaches can ambulate their limbs at astonishing speed while having barely a few thousand neurons for brains, there must be something else to how they do it.
He went to work building distributed sensation into his robots, allowing the limbs to meter the degrees of freedom and thus reduce the complexity of the ambulation calculations....however the real innovation came when the application of statistical learning algorithms combined with these distributed sensors on limbs made it's debut.
These methods allow for a massive collapse in computational requirements by simply training the limbs to "replay" previously stored successful movements for a priori sensed positions of body and limb positioning. This allows the robot to "remember how to walk" rather than "computer how to walk" for every ambulation cycle. In the early 2000's many teams applied genetic programming to train mechanical robots in virtual environments ...allowing them to build the statistical maps of successful ambulation for given terrain encountered....and here we are...2013 and seeing the fruit of these advanced methods in the work of Raibert (now head of Boston Dynamics).
Big Dog
PetMan
Little Dog
Atlas
:All use these critical insights of distributed sensation and statistical learning to reduce computational complexity by orders of magnitude BUT all the proof we need to know that current method still have lots wrong is that no robots currently are as fast or agile on uneven or mixed terrain as a roach.
To me this means quite simply that either the type or amount of distributed sensation necessary to achieve that level of dexterity is not yet discovered. I don't think it has anything to do with the computational muscle...which at this point is quite overkill for the problem by orders of magnitude.
The work I've been doing with the Action Oriented Workflow algorithm is very much related to these ideas, AOW is a generalized algorithm for defining arbitrary "action" attributes these are then sampled in as dense or sparse a set as necessary to gather data on when those actions are performed. The Action Delta Assessment algorithm is the underlying statistical learning function for the algorithm, allowing historical information for action execution to be compared in real time in a distributed fashion...precisely as what is needed to refine ambulation in robotic limbs. It is quite simple to see that mechanical ambulation of limbs is directly analogous to this , I have some ideas on how I can apply the algorithm to discover and learn these patterns but am focused on applying AOW to the abstract action space of interacting human and business objects in software (to learn and refine). The fractal nature of the algorithm makes it ideal for solving problems of very messy data sets once the sufficient level of resolution to the "action" points are made.
Links:
http://sent2null.blogspot.com/2009/04/agilentity-architecture-action-oriented.html
http://en.wikipedia.org/wiki/Statistical_learning_theory
http://www.bostondynamics.com/
http://en.wikipedia.org/wiki/Marc_Raibert
http://sent2null.blogspot.com/2012/02/with-completion-of-ada-action-delta.html
Comments