Neural Network Models for Skills
The network of neurons of the brain not only helps us perceive the world, but also guides our behavior in that world.
------------------------------------------------------------------------------
--------------------------------------------------
The intricate neuronal network we call our brain lets us perceive, or form a virtual reality we call the phenomenal world. It also provides for the synchronous contraction or relaxation we need for each and every one of our muscles in order to guide our movements. So we have the means to develop skills, which I view as learned sensory-motor interactions.
We somehow learn to make distinctions to create perceptions and recognize sensory patterns; this is in our nature as perceptual beings. Through bumbling experience we begin as babies (in a symbiotic relationship with our caretaker mother) to develop the know-how (the conceptual framework), to sequence our muscles to behave in appropriate, understandable ways (behavioral patterns), in what we call learned or skilled behavior.
Coordination
of behavior (patterned motion, or know-how) with perception (patterned sensing, pattern recognition, information acquisition, situation awareness -- know-what) requires that the brain perform appropriate transformations of the sensory components into motor components and motor components into sensory components -- a complex, highly interactive read and react principle. Inasmuch as know-how and know-what interact and are extended in time, the brain must generate an ongoing sequence of motor vectors whose changes over time produce the right behavioral changes.The transformations have to be based on the configuration of the brain's synaptic weights. This is where intelligence seems to begin, which is to say doing the right thing at the right time in its perceived environment. This is decision-making or self-mangement in its most intimate form. It is also where skills hang out, where smarts are located.
The question arises whether artificial neural networks can be used to study these sensory-motor interactions. Here we try to approach the study of skills through the use of these artificial neural networks. There is no fast-track for the study. Nor is it clear that neural networks, alone, can even get close to representing skills, inasmuch as understanding is an essential characteristic of skills and may not be, or ever become, part of any neural network.
There is this, however: Construction of a net requires inputs. And inputs are organized or patterned elements. Thus they already involve an organizer, an intelligent entity with modeling capability. So they already involve skills. The network as such would then learn to take the input primitives to a higher conceptual level via the intermediate layers of connectivity.
--------------------------------------------
What is it that has to be done with neural networks? What exactly is the project as far as, say, the stock market or tennis is concerned?
The task as I see it, is to find a way to represent the events taking place in the sensory-motor interactions that make up our experience. That's a tall order, because experience involves observing what goes on in some environment and acting on what is perceived. And that could be just about anything. Interactive sensory-motor skills are involved in purposive behavior. And the skills could relate to physical, psychological, or social activity.
In the stock market it means discerning the state of the market and buying or selling some equity based on what you observe. In tennis, it means tracking an oncoming ball and racing to intercept it and smashing it back to your opponent. But this is only very general stuff. Each activity is very complex. To do something meaningful, we have to get down to the nitty-gritty of the processes. That means developing "neuronal" clusters to handle specific aspects of each of the problem situations, which means developing a neural network practically as complex as our brain.
Perception
Lets first look at the tennis problem, because the playing arena (the observation context) is limited to the court and so the observation procedure is more clearly outlined than in the stock market environment, which could extend to the whole world. So suppose you are the tennis player and you are standing at your end of the court watching and waiting for your opponent to serve. What does this imply?
For one thing, it means you already have a way of distinguishing objects and recognizing what's occurring, meaning that you recognize you are playing tennis against an opponent in a normal tennis arena. Specifically, it means you can detect and identify the object we know as a ball (i.e., make a distinction between the ball and the other stuff). It also means you can, for example, follow the ball as it:
The reason for watching the ball is to get information about it -- information like where it is at the moment, where it is going, how fast it's moving, and so on. You acquire the information as a recognized pattern.
However, even if you're a good athlete, it's unlikely you'll be able to track the ball effectively and intercept it for a return if you've had little or no playing time in tennis. I've taught novice tennis players who were otherwise good athletes and all had difficulties reading the trajectory (recognizing the pattern) well enough to intercept the ball. While you might have no trouble just seeing the ball move, it's another matter entirely when it comes to being able to tell (in body-movement terms) how fast it's going, predict quickly where it will go, know how it will bounce, and determine where you might best intercept it. It may move faster than you expect, and might bounce higher. So you can easily miss-hit the ball or be late and have it get by you. Hitting the ball correctly is acquired know-how. Practice and a lot of playing time are needed to develop an understanding of the motion. This is also a critical requirement for a good neural network model and makes you wonder if a simulation using nets can be meaningful.
In the stock market, perception means keeping track of the price movement of a stock (or an index of some sort), getting information about it (like volume, the range of movement, closing price) and predicting where the price will go next. There's no obvious physical relationship here as there is in tennis. The price definitely doesn't move through physical space, though it does "move" across the chart if you happen to be plotting it, and it definitely "moves price-wise" in time. Here, too, to make sense of what's going on you have to develop an expertise. You have to acquire know-how. In other words, you have to be able to read corporate behavior or read the charts (price behavior). This is also what we want the neural network to develop. Needless to say, much learning is involved here, too -- on my part as well as that of the network.
Response
As a receiver of the tennis ball, you have to respond to it by rotating your body and pulling your racket back in preparation for a swing, and likely shifting laterally to be in position to make effective contact with the ball. Assuming you know where the ball is going -- i.e., that you recognize its trajectory -- you still have to direct your body to the correct or best interception point and swing the racket in the appropriate way to make a return shot. This requires much experience and a great deal of learning, as well as conditioning, because it means sequencing the contraction and relaxation of all of the muscles in your body in the proper order to be able to reach the ball and make the return.
The stock market is no different in principle from tennis as far as response is concerned -- certainly not in the sense that the response has to be timely. The movements you would have to make are, of course, completely different from those in tennis. You have no racket to wield, and you don't smash a ball. But you still have to do things like placing an order to buy or sell shares, and you should best do that at the right moment, to be in "position" to benefit from the anticipated "move." You want to "intercept" the price trajectory at the right time, i.e., when it has reached the right value. To do this effectively, you may have to be "fast on your feet." And you sure need to understand what's going on.
Interaction
For a high speed serve your response to the ball is almost instantaneous. Or so it seems if you have to hit it. Whether the motion is high-speed or slow, however, you till have to detect where the ball is going, and you have to adjust your position and attack posture to meet it. The adjustment changes your perspective a tad and you obtain more information about the moving ball from your new viewpoint. You might then change your position once more and adjust your racket status again, and so on, repeatedly, until you either hit the ball or it passes by you. This is the case whether you have only a split second to react or have time for a cup of tea before the ball arrives.
The interaction in regard to the stock market is different from tennis in that you don't particularly change your position to adjust to the price movement. For instance, you don't move closer to the phone. Even so, with each price change you do get a new perspective of the stock and this may lead you psychologically "closer" to buying or selling. However, the change in your perspective has no effect on the price, even though your own attitude might be changing. But neither does your perception of the tennis ball affect the trajectory of the ball. So there you are! There are no obvious quantum effects here.
----------------------------------------------