Next: Making the Net Work
Up: Table of contents
Previous: The Social life of computers

Bottom-up design of self-organized complex systems

In spite of this analogy with previous complex systems, the systems made of man-designed elements offer an unprecedented opportunity. While the complexity emergence in biology and society took place unintentionally, often as a result of chance, in man-made objects the emergence can be planned and influenced through the way in which the elementary objects are designed and displayed. The convergence of complexity and IT paradigms enables for the first time the deliberate development of a new breed of complex applications and systems:

Systems designed so that the simple interactions between their elementary components create collectively a desired global behavior.

Using the methodologies developed to study complexity phenomena, one can design complex behavior in artificial systems. This paradigm has the potential of transforming entire fields of IT applications. It may boost the development of directions that otherwise would reach a dead-lock, or would be increasingly limited by a mere logarithmic growth with the computer power. For certain applications, e.g. network routing, orthodox, global, hierarchical design methodologies can be proved to be inapplicable, even if the hardware capabilities will continue to grow exponentially. Similar examples of such fields and applications are:

• Car traffic control optimization
• Texts translation
• Information retrieval and knowledge extraction
• Network routing, content routes optimization
• Adaptive collaborative software development
• Real time air traffic scheduling adaptation
• Autonomous UAV (Unmanned Aerial Vehicle) flocks
• Bioinformatic research algorithms and tools optimization
• Integrated robot flocks (or 'dividuals', see below)

Below, we look further into some of these examples, but first let us discuss the generic advantages and problems of such 'bottom-up' designing of complex systems. The most important are their

1. freedom of scale
2. adaptability and robustness
3. ease in design, implementation and maintenance

In algorithms theory the O(N) computational complexity measure is the corner stone of the entire science. The functioning of many of the 'bottom-up' designed systems is often 'simple' in this respect: all of the work is done locally and the issue of the cost at the level of the entire system does not even arise (nobody has yet submitted the next year budget for the world web :-). Similarly, one may add autonomously traffic lights, install new routers, or add additional UAVs to a flock as much as one wants, and the robustness or cost of the rest of the system will not be significantly affected. The local adaptive-agent based control algorithms become unavoidable in applications where in principle there are possible interactions between any arbitrary elements of the system. In such examples, the global hierarchical algorithms have to deal with an exponentiation in the number of alternatives to be considered. The local adaptive agents are usually dealing well with avoiding to even considering the cases that do not actually appear during their current functioning. Usually this is a great relieve as most of the computationally 'worst cases' never appear in practice.

The high non-linearity of certain systems, like the control of traffic lights in a large city, makes it virtually impossible to find the optimized global algorithm for these systems. However, using methodologies developed in the study of complex phenomena optimality and predictability become possible, and in addition the systems themselves become understandable.

It is extremely hard, and in many cases plain impossible, to built a global logic system with good reflexes. By good reflexes we mean the ability to rapidly adapt to changes in an adequate manner. The need to tell in a top-down manner to each element what to do next makes it extremely difficult to react fast in the correct way. However, when one gets a collective behavior as an emergent character of a multitude of elements, adaptation comes naturally, and only in regions where it is needed.

Another important advantage of 'designing for emergence' is the fact that it makes the applications easy to design, program and maintain. This is crucial for keeping-up the expansion rate of computer uses in domains with low programming literacy. There are simply not enough programmers around to design, implement and maintain classical software tools for every user.

The big disadvantage, but truly the big challenge, is that there exists no standard theory and practice for designing systems that emerge self-organized complexity bottom-up. It is obvious that this will require not only technological progress but also scientific effort. On the other hand its results might fare much beyond efficiency and impinge upon the very frontiers of human knowledge.

Let us look now on several promising examples of IT applications that may benefit from applying complexity concepts.

Specific Applications

Agent based and collaborative applications

There are two approaches, about agents cooperating in the management of complex tasks. The more traditional approach is making the agents more and more sophisticated, with AI core etc. The second and more promising approach aims 'collaborative applications' i.e. applications that encompass multiple agents in a collaborative manner in order to fulfill a task.

The main conceptual limitation of the current collaborative application approach is that it tries to divide the global task into small independent segments performed by independent agents. This external division and stream-lining of the tasks preserves some of the rigidity and weaknesses of the logical-tree, centralized methods. The challenge is to design the right interaction protocols and feedback mechanisms that to insure the self-organization of the work in an optimal way.

Traffic Lights on the Spot

Centralization was a major paradigm for control theory since its inception. And for good reasons. A centralized control system is deterministic, and can be developed in a straightforward manner. But currently we are in a convergence point of several processes that suggest a coming paradigm shift. First of all, the amount of computer power which can be squeezed into a portable computerized device has grown exponentially. Even more important is the exponential growth in communication abilities, which makes both land based and wireless communication technology a mature, cheap and accessible option. These technological developments are met with new research paradigms described above, aiming to study the complex systems behavior that emerges from the interaction of the system's components. Among control applications, traffic management looks especially promising. It is composed of many identical components, which need to adapt rapidly to changing circumstances. A centralized system, wishing to adapt to a car accident will need to check all of the global ramifications of a change, while in a local-rules-built system such a process may occur naturally.

Ultimately, the very concept of traffic light may come under revision. The same information, instructions, regulations and signals which are currently communicated through road signs might be transmitted and may be even enforced at the level of the individual cars: Your car will slow and stop at 'red lights' unless you explicitly choose to override the order it receives from the 'traffic lights' system. Reciprocally, the traffic regulator program will take into account your travel plans, constraints, and the condition of your car in issuing its orders to the other cars.

Self-Organized Unmanned Aerial Traffic

The concept introduced above: emergence of self-organized traffic is particularly fit for Unmanned Aerial Vehicles (UAV) navigation:

• in aerial traffic there are less intricate obstacles than in ground transportation,
• the danger for hurting humans is less direct in unmanned vehicles
• humans do not have a head start in 3D navigation skills compared to computers (as opposed to hundreds of thousands of years of 2D navigation).
• The 3D geometry allows for the formation of large flocks that form, join and split according to the individuals initial and final destination points.

Relaxing the demand for a central control center that communicates with all of the UAVs in real time and directs them what to do, will enable creating large flocks of UAVs, that will be able to travel much further from their home. In addition it will enable to enlarge significantly the set of possible tasks such flocks will undertake, due to the fact that their reaction time scale will be reduced dramatically. In the case of Cruise Missiles Flocks the possibility to share their location and visual information, may result in a dramatic improvement of their navigation and target identification skills.

From Integrated Robot Flocks to Dividuals

The idea that a collection of objects sharing information can be more efficient than a more intelligent single object has tremendous potential.

As opposed to humans, robots can share and integrate directly visual, and other non-linearly structured information. They do not suffer from the humans need to first transform the information in a sequence of words. Moreover, the amount, speed and precision of the data they can share are virtually unlimited.

Like in the story of the blinds feeling an elephant, the communication channels typical to humans are not sufficient to insure fast, precise and efficient integration of their knowledge / information / intelligence. By contrast, robots with their capability to determine exactly their relative position and to transmit in detail the raw data 'they see' are perfectly fit for the job perfectly.

So, in such tasks, while

1 human > 1 robot

one may have

100 humans < 100 robots.

This may lead to the concept of Integrated Robot Flocks which is much more powerful than the biologically inspired ants nest metaphor because the bandwidth of information sharing is much more massive. Rather than learning from biology, we may learn here how to avoid its limitations.

For instance, rather than thinking of the communicating robots as an integrated flock, one can break with the biology (and semantics) inspiration and think in terms of the divided individual (should one call them 'dividuals'?):

Unlike biological creatures, the artificial ones do not have to be spatially connected: it may be a great advantage to have a lot of eyes and ears spread over the entire hunting field. Moreover, one does not need to carry-over the reproduction organs when the teeth (mounted on legs) go to kill the pray (the stomach too can be brought-in only later on, in case there is killing).

Of course, a good idea is to steal from time to time the control on somebody's else wings (as a form of Non-Darwinian evolution). Contrast it to the usual way real animals exploit one another: they are constrained by their biological reality to first degrade the wings of the prey to simple molecules at which stage, it is too late to use them for flying.

The anthropomorphic- biological grounding here is a liability that one has to free oneself from, rather than a source of creative inspiration. All said above is true both for robots (acting in the real physical 'hardware' world) as well as for 'bots' acting as software creatures.

Encounters of the Web kind

The emergence of a 'thinking brain' by the extension of a distributed computerized system to an entire planet is a recurring motif in science-fiction stories and as such a bit awkward for scientific consideration. Yet, if we believe that a large enough collection of strongly interacting elements can produce more than their sum, one should consider seriously the capabilities of the web to develop emergent properties much beyond the cognitive capabilities of its components. As in the case of the Integrated Robots Flocks, the relative disadvantage of the individual computer vs. the individual human is largely compensated by the 'parapsychological' properties of the computers: any image perceived by one of them at one location of the planet can be immediately shared as such by all. Moreover they can share their internal state with a precision and candor that even married couples of humans can only envy.

A serious obstacle in recognizing the collective features emerging in the web is the psychological one: people have a long history of insensitivity to even slightly different forms of 'intelligence'. In fact various ethnic / racial groups have repeatedly denied one another such capabilities in the past. Instead of trying to force upon the computers the human version of intelligence (as tried unsuccessfully for 30 years by AI), one should be more receptive to the kind of intelligence the collections of computer artifacts are 'trying' to emerge.

An useful attitude is to approach the contact with web in the same way we would approach a contact with a extraterrestrial potentially intelligent being. A complementary attitude is to study the collective activity of the web from a cognitive point of view, even to the level of drawing inspiration from known psychological processes and structures.

Next: Making the Net Work
Up: Table of contents
Previous: The Social life of computers