Chapter 14 3.4 Advantages of nested hierarchies

Inside every single living entity is a whole host of non-living things.Someday there will be a whole host of non-mechanical things inside every single machine.No matter what type of group they are, on the one hand, they are all busy in their own way, and on the other hand, they form a new whole. Brooks wrote: "The subsumption architecture is essentially a parallel distributed computing that connects the robot's sensors and actuators." The point of this architecture is to break down complex functions into small unit modules and organize them in a hierarchical form.Many observers relish the social ideal of distributed control, but are disgusted to hear that hierarchy is the most important and core part of the structure of inclusion.Doesn't distributed control, they ask, mean the end of hierarchy?

As Dante ascends the tiers of the empyreans, he climbs a hierarchy of status.In the status hierarchy, information and power flow in one direction from the top down.In an inclusive or networked hierarchy, information and power flow from the bottom up, or from one side to the other.Brooks noted that "regardless of which layer an agent or module works at, they are all created equal... each module just has to keep its head down and do its own thing." In the distributed control management system of human beings, certain types of hierarchies will be strengthened rather than diminished and disappeared.This is especially true in distributed control systems that involve human nodes -- such as vast global computer networks.Many computing activists are touting a new era of network economics—one built around peer-to-peer networks of computers—and arguing that it's time to move away from those hierarchical networks.They are both right and wrong.Although the authoritarian "top-down" hierarchical structure will tend to die out, distributed systems will not last long without the nested hierarchy of "bottom-up" control.When individuals in the same layer interact with each other, they naturally aggregate together to form complete cellular organs and become the basic units of larger but slower-moving networks.Over time, a multi-tiered organization based on seepage control from the bottom up develops: faster activity at the bottom, slower activity at the top.

A second important aspect of generic distributed control is that the sub-aggregation of controls must be incrementally accumulative from the bottom.It is impossible to disassemble complex problems into logical and interacting factors through reasoning.Good intentions are bound to fail.For example, some large and improper companies in joint ventures have a very high probability of collapse; large institutions created to solve the problems of another department become themselves problematic departments. In mathematical operations, division is more difficult than multiplication, and for the same reason, top-down classification aggregation is not feasible.It is easy to multiply several prime numbers to get the answer, and elementary school students can do it.But to decompose a large number into a prime factor, the most supercomputer will get stuck.Top-down control is as difficult as factoring a product into factors that are very easy to factor.

The relevant laws can be stated succinctly as follows: distributed control must be derived from simple local control; complex systems must be derived from simple systems that already exist and work well. To test the bottom-up distributed control theory, University of Rochester graduate student Brian Shanuchi built a robotic arm called a juggling toss.The task of the arms is to repeatedly tap a balloon with a paddle.The robot arm does not have a brain to position the balloon and direct the beat to move under the balloon, and then use the appropriate force to flick the beat; instead, Yamauchi decentralized the work of positioning and controlling the force.The final balance of motion is done by a committee of dumb "agents".

For example, subdivide the most complex puzzle of "Where's the balloon?" into several separate problems, spreading it across many tiny logic circuits.An agent considers only one simple question: Is the balloon within reach? ——A relatively easy-to-operate question.The agent in charge of this problem has no idea when the balloon was slapped, or even where the balloon is.Its single duty is to instruct the arm to back up when the balloon is out of view of the camera on the arm, and to keep moving until the balloon comes into view.A network or society of these simple-minded decision centers forms an organism capable of displaying remarkable agility and adaptability.

“There is no explicit exchange of information between behaving agents,” Yamauchi said. “All communication occurs by observing the traces and effects of other agents’ actions in the external environment.” Keeping things local and Immediateness allows society to evolve new behaviors while avoiding the complexity explosion that accompanies the "hardware" communication process.Contrary to popular business preaching, telling everyone everything is not how wisdom is made. "We took this idea a step further," says Brooks, "and often use the outside world as a communication medium between distributed parts." Instead of being told what to do by another module, a reflective module senses the outside world directly. The reflected information is then transmitted to others through its effect on the external world. "It's possible for messages to get lost -- quite often actually. But that's okay because the agent keeps sending messages over and over again. It keeps repeating 'I saw it. I saw it. I saw it' message , the agent will not be quiet until the arm receives the information and acts accordingly to change the external world."