Governing Agents

The intelligent leading the intelligent

Awareness of the information-processing power that genetic-based evolution represents has led many people to wonder how we can harness that power for other information-intensive endeavors. We want to apply techniques learned from natural evolution to business research, decision support, knowledge management, financial-instrument trading strategies, software development, and medical research. We would even like to apply it to understanding genetics itself — unraveling the implications of the human genome for health and illness, in the new field called bioinformatics.

People have applied genetic algorithms to such diverse areas as setting connection values in neural nets and designing walking robots. But the introduction of the agent concept and extensions to that concept, such as architectures of large systems of agents (ALSAs) give us a far more elaborate conceptualization of what DNA accomplishes in terms of information processing. Empowered with this new understanding, we can create better applications in other domains.

In “When the Agents Come Marching In” (my July 17 column), I provided an introduction to ALSAs, suggested four main kinds of architectures, and described the potential for businesses to be run like well-orchestrated marching bands. In my last column (“The Wisdom of Critters and the Madness of Crowds,” Aug. 1), I elaborated on this Muddy Waters architecture, in which agents (smart things that perceive their environment, make decisions, and then take actions) interact only if they happen to be proximal in space, time, or information flow. Both ant colonies and markets seem to follow this structure. Ant colonies are highly adaptable and capable of routing around congested pathways or finding new resources. Their computer counterparts are thus highly useful for many kinds of scheduling and routing problems, including network traffic and industrial production scheduling.

The ant colony’s results stand, however, in stark contrast to open markets’ results. Markets sometimes manifest “irrational exuberance,” as Federal Reserve Chairman Alan Greenspan put it, such as stock-market extremes and the consequent potential bubbles and crashes. Do we really want our business-supporting systems to be built of smart components that just bump around in the dark hoping to get the information they need by happenstance within unpredictable system dynamics? No, of course not. What’s the difference between the ant colony and the open market? Governance. The colony is a governed ALSA (GALSA).

Organized marching bands are highly governed by marching scripts, music scores, and a conductor. The ant system owes its success to a more hidden form of governance: ants’ common DNA. You can say that DNA sets the general rules for ant-colony marching orders, with these instructions evolving toward more successful system results over time. The regulations you may find in markets just don’t compare to DNA’s level of governance. Let us look first at how DNA governs agent systems and then compare it to processes occurring in natural language, in order to apply these more general information-processing approaches elsewhere.

DNA synthesizes proteins (mostly), and each protein, because of its chemistry and geometry, can interact with some of the other proteins in the biological soup. Depending on conditions, these components can produce new products or modulate other processes. (See Figure 1) These proteins fit the agent definition — they perceive the environment and act in response to conditions. Biological pathways can be seen as sequences of operations in these agent interactions. The governing DNA is an agent also. It reacts to a specific context — whether a cell is in the kidney, brain, developing embryo, or part of a genetically induced cancer — and generates the proportion of proteins and the pathway modulators. It thus defines the balance of biological pathways chosen (among potentially a great many) and what in the end gets synthesized or constructed (such as a cell membrane or a neural transmitter).

FIGURE 1 Biological pathways can be seen as analogous to language, giving us a model for GALSAs acting on text.



Click Here for PDF Version of figure

The protein agents are not entirely under the control of DNA, which only gives a constraining set of options or higher-level instructions. The agents and the generated system itself respond to different conditions such as adrenaline presence in the blood. This relationship could be compared to higher-level business management giving general orders to experienced company negotiators, such as, “Do whatever it takes to close the deal.”

What is interesting is there are many levels of context-dependency. The governing agents react differently in different contexts, and so do the governed agents — just at a lower, more specific level and more quickly.

These biological pathways are subject to at least two kinds of constraints: the raw potential of chemical and geometric options, and the ranking of those options via the availability of agents and process modulators.

The agent-governing structures I described earlier have a remarkable parallel in the information-processing solutions found in natural language systems.

Biological pathways parallel the sentences spoken in natural languages: They are generated in real time, defined by systemic constants both universal and particular to a given context (here language, sublanguage, and discourse situation), and they have specific, context-dependent functions.

Work by Paul Smolensky at the Cognitive Science Department and the Center for Language and Speech Processing at Johns Hopkins University suggests that the universal or innate grammar humans share is really a list of general rules. These rules parallel the general rules of chemistry and geometry that are the first-order constraints in biological pathways and include such mandates as, “All words have a use,” “All sentences have a subject,” and so forth. The grammar of a specific language is the governing agent that gives a precedent order for these rules, for situations where they conflict. (He has similar arguments for the phonetic and semantic levels.) Professor Smolensky has shown that both children and adults learn languages far faster when presented with grammar in these terms, which is a good indication the theory actually matches the processes our minds use.

The usual, simple way of comparing DNA and language is to say DNA’s four bases (represented by A, C, T, and G) are like letters in an alphabet and sequences of these are like words and sentences. However, DNA is a comparatively static but generative structure and is thus more comparable to language as formal structure — something you can investigate at the level of a population as a whole.

DNA generates the “vocabulary,” the set of available proteins useful in that context, and the pathway modulators, analogous to grammatical rules, which rank the dominance order of the universal (chemical) rules. It is the biological pathways that have the properties of sentences as generated in time. You could even extend the analogy to the “discourse” between cells.

In both cases, what we have are very compact ways of defining a set of instructions that will guide complex system elaboration in interaction with potentially wide ranging sets of environments. Governing agents can generate, facilitate, and control the sets of agents, resources, and constraints already available in an environment to create anything, from a mouse to a Maltese.

When it comes to DNA and languages, these governing agents can also evolve to increase system adaptivity.

There are enormous implications in all these features of governance for potential use in information processing and engineering successful context-sensitive systems. Each combination defines a different kind of governance and deserves careful examination.

The complexities of creating compact codes that generate agents so that there is a continuous map between the code and its manifestations, as with DNA, is an additional feature that is far beyond what I can cover here.

I can only list some examples of applications of one combination of these features — genetically varying governing agents that choose agents and that facilitate or control agent-to-agent or agent-to-resource interactions. These could include several options for agents used and their connective syntax, information connectivity, or rule dominance. These kinds of genetic searches are useful when the search space of potential agent systems is prohibitively high yet finding constantly improving solutions is tremendously valuable. This is certainly something far beyond just the varying of parameters in the ways current genetic algorithms do to find neural net solutions.

In Table 1, I list some of these applications. Here I focus on agent governance in its potential role as offering a genetic search of system options so as to improve system dynamics.

TABLE 1 Examples of application domains where governing agents can be genetically varied to select subsets of potential agent types and rules to discover ever-improving solutions.

Agent types selected from	Rules	Architecture*	Application
1. Signal transformation Neural nets	Options for ordering syntax	FA	Discovery of trading rules
2. Image transformation Neural nets	Options for ordering syntax	FA	Face and image recognition
3. Kinds of trading	Regulations	MW	Trading-regulation testing for market pricing stability
4. Network administration	Order of precedent of actions	AO	System dynamic stability
5. Business rules	What-if and if-then scenarios	FA	Systematically working out multileveled strategies
6. Business rules Orchestrating rules	Hierarchy of precedents Points of contact	MW	Systemizing business processes
7. Collaborative filtering rules and options	Ordering and organizing over a domain	AO	Finding better ways to integrate the knowledge of many participants
8. Unstructured search, selection, transformation, ranking, and reduction	Options for ordering and syntax	FA or AO	Creating methods for finding previously unknown facts in large text databases
9. Method for extracting meaningful genomic sequences Mapping operators to proteins associated with diseases	Sequence of operations	FA	Discovering gene and disease relations
*The architecture the table references are the ALSA types outlined in “When the Agents Come Marching In.” MW=Muddy Waters, FA=Formal Agents, AO=Agents at Outposts.

A separate topic that deserves extensive extrapolation is the set of options for using hierarchies of governing agents as a means to manage large systems of agents. The idea is to create excellently managed systems with intelligent units managing still other units with each level handling a different level of detail.

Certainly the governing-agent paradigm, which seems to show that the tree of life is simply the tree of knowledge by another name, is an amazingly rich one to explore.

Sidebar

EVOLUTIONARY MARKETING