A multi-celled organism, anything from a small nematode to a large primate, consists of specialized cells organized into a variety of communities. In a primate, for example, there are organs like the heart, lungs, various muscle sets, brain, and so on, each specialized to carry out a distinct activity such as pumping blood, adding oxygen to the blood, manipulating the environment, etc. Although counterparts of all these activities can be found in a single free-living cell, such as an amoeba, communities of specialized cells allow more efficient execution of the activities (as we’ll see shortly).
Community-based organization extends both downward and upward from the level of organisms, giving rise to the hierarchies that characterize complex adaptive systems (CAS). For example, an individual biological cell is delimited by an outer membrane; within that membrane are further membranes delimiting organelles, the nucleus, etc. The nucleus in turn contains clusters of proteins centred on DNA. All this apparatus greatly enhances the manipulation of the cell’s basic building blocks, proteins. In the upward direction, there are ecosystems where the term niche is used to describe communities of interacting species (more about this in Chapter 7), and on up to continent-wide clusters of ecosystems.
When one looks at biological communities, it is amazing just how specific some interactions can be. Darwin’s ‘comet orchid’ provides an example that also shows that the theory of natural selection can be used to make strong predictions. Darwin made a careful study of orchids and knew that orchids related to the comet orchid are pollinated only by moths. Because the nectar tube of the comet orchid is a foot long (hence the name), Darwin surmised that there must be a moth with a foot-long proboscis (!) that specializes in taking nectar from (and pollinating) the comet orchid. The moth involved, a variety of hawk moth, was finally filmed over a century after Darwin had made this prediction. Because of their extreme co-evolutionary adaptation, neither the moth nor the orchid could survive without the other. Counterbalancing this weakness is the fact that the moth has no competitors for its food source, and the orchid wastes no nectar on inefficient pollinators.
Communities provide for the interaction of specialists, but what advantage does this interaction have over the independent activity of an equivalent number of generalists? Adam Smith’s famous example provides an important clue. Just prior to the time Adam Smith was writing, straight pins were produced by blacksmiths (generalist craftsmen) through a combination of skills: drawing wire from molten metal, clipping the wire to produce a sharp point, and adding a blunt head to the other end of the wire. The process was difficult and time-consuming, making straight pins a luxury item. Then, at the time Adam Smith was writing, one of the first ‘production lines’ formed for the purpose of producing straight pins, each step in the process being carried out by a specialist. Throughput increased by a factor of 10, and straight pins became widely available.
Similar increases in throughput are observed when one looks at membrane-separated cascades of interactions within a biological cell. The semi-permeable membranes enclosing organelles admit some proteins and not others. The immediate effect is an increase in the concentration of the admitted proteins within the organelle. By the laws of elementary chemistry, the higher concentrations increase the reactions between the admitted proteins. Other organelles in the cascade can then process the products of these reactions, yielding a production-line-like cascade of specialists, à la Adam Smith. The result is a substantial increase in the availability of products involved in the cell’s survival and replication, thereby increasing the cell’s Darwinian fitness.
All CAS that have been examined closely exhibit trends toward increasing numbers of specialists. As an example, consider the progression of changes from the market squares of early towns to present-day commodity markets. In the early markets, individuals exchanged self-produced products, such as hand-woven cloth, eggs, leather, and so on. A present-day commodity market consists of a vast array of specialists dealing in futures, hedges, derivatives, etc., none of which involves handling the actual commodity. Looking to other CAS, the rampant specializations in contemporary automotive production lines or government bureaus are a far cry from their earliest versions. Similar observations apply to the Internet, weather bureaus, and flight control, to name a few others.
In every case, there are diverse specialist agents that attend to, and process, selected signals. As in biological cells, the diversity is sustained by a combination of boundaries (the cell’s semi-permeable membranes), signals (the proteins), and signal-processing (the reactions between proteins). A surprising characteristic of this signal-processing is that relatively small parts of signals, which I will call tags, route the signals through the boundaries. Tags have different names in different CAS studies—‘active sites’ (proteins), message headers (the Internet), ‘motifs’ (molecular genetics), and grammatical declensions (languages), and so on—but the ‘routing’ functions are similar. For example, small parts of protein signals cause them to adhere to specific parts of chromosomal DNA, turning genes ‘on’ and ‘off’ and enabling complex signal-dependent computations. This tag-based control allows cells to exist in a diverse range of conditions by activating only genes that are relevant to the current situation. Similar advantageous relations between tags and boundaries can be found in CAS of all kinds.
Tags have a particularly important role when they are used to coordinate the rule sequences exemplified by cascades of reactions, ‘production lines’, and the ‘middleman’ sequences discussed under credit assignment in Chapter 3. When tags coordinate ‘production lines’, variations with substantial differences in throughput and efficiency are easily generated and tested. Because many kinds of production-line-like cascades are possible, CAS agents can develop ever more complex strategies for survival. And each new kind of agent offers still further opportunities for interaction. In this increasing set of possibilities for interaction we have the beginnings of an explanation for the pervasive diversity of CAS—new specializations open the way for still further specializations and still greater throughput and diversity.
To get a more precise view of the routing possibilities for tags, let’s look again at the use of #s to define the conditions for signal-processing rules (introduced at the beginning of Chapter 4). A condition can be set to respond to a given tag by using #s (don’t cares) on either side of the tag; thus the condition #100### … # responds to a signal string with tag 100 at the 2nd, 3rd, and 4th positions of the string. Note that a condition with many #s can accept a variety of tags while a condition with few #s will accept few if any tags (requiring a specific string when it has no #s). Or, looking at tags themselves, short tags satisfy a variety of conditions, while long tags make highly specific requirements on conditions.
In a binary coding, only a few positions are required to provide for hundreds of distinct tags (ten positions suffice for over 1,000 distinct tags). Thus, tag diversity is easily achieved, and relevant routings are easily maintained. Moreover, interactions can be simply re-directed by making modest changes to tags. The simple relation between generality and specificity under the # notation offers a direct way of transforming rules, and the agents containing them, from generalists to specialists, or vice-versa. The question then becomes: How does a CAS discover or generate relevant tags and the sequential routing that yields ‘production lines’?
As just pointed out, it is simple to modify tags, and the conditions that route the tags through boundaries, to generate a wide range of interactions. The problem is to do this in ways that build on an agent’s experience. Cross-breeding, discussed earlier in connection with rules (under ‘Rule discovery’ in Chapter 3), offers a good way to accomplish experience-based generation of new tags. Consider two rules that are crossed in the tag-region of their condition (signal-receiving) part. Let the two rules each accept a distinct tag, and let the tags occupy overlapping loci in the string, with one condition requiring a more specific tag (more non-#s) than the other. The result of a cross within the region of overlap between the tags will be two new rules, with condition parts requiring new tags (as a result of the exchange of parts). Generally, one of the condition parts accepts a wider range of tags than either parent (it has more #s in the tag region), while the other is more specific (see Figure 8).
A rule-condition with a more general tag requirement can accept signals in addition to those accepted by either parent. In particular, this new rule may accept signals from another production line. The new rule becomes a new buyer for some rule in the other production line, thereby supplementing that production line’s income. If the new rule sets the stage for a new use of the original production line, and the resulting throughput is valuable, then both the production line and the new rule will prosper. Under the bucket brigade credit-assignment procedure (‘Credit assignment’ in Chapter 3) there will be corresponding increases in the strength of all the participating rules, thereby assuring the survival of the augmented system.
The same procedure can generate new production lines ‘from scratch’. Consider a rule that often gets immediate rewards but has no coupled supplier, so that it is often ‘unprepared’ when reward opportunities occur. Then, let cross-over generate a rule with a signal (resource) that satisfies the condition of the rewarded rule. The newly generated rule thus ‘sets the stage’ for activation of the reward-acquiring rule, making it more likely to be rewarded when activated. When this happens, the coupled pair acts as a two-agent production line. The two-agent line then becomes a candidate for further stage-setting, providing opportunities for still longer lines.
Generally, when the output of a production line is valuable it is used up rapidly—as when an agent sequesters the output to help it reproduce. If we look upon the production line as a sequence of chemical-like reactions, then ‘valuable’ end-products of the sequence are rapidly sequestered, lowering their ambient concentration. The resulting lower concentration of end-products lowers the back-reaction rate (disassembly)—much as in Herb Simon’s classic comparison of two watchmakers. Thus, under the laws of elementary chemistry, there is an increase in net (forward) throughput of reactions leading to valuable, sequestered end-products. In other words, production lines with valuable outputs are more efficient at acquiring and processing relevant reactants, so they tend to dominate an agent’s activities.
In CAS agents, many rules and production lines are active simultaneously (recall the activities in a biological cell). Because new rules generated by cross-over do not usually replace their parents (replacing weak rules instead), extant production lines are rarely disrupted by new rules, allowing exploration without disrupting regularities already being exploited. Under this procedure for generating new rules, tags serve as building blocks that can be recombined to foster new production lines. The next chapter will take a closer look at the relations between building blocks, generated systems, and the phenomenon of emergence.