Some thoughts on multi-agent systems and Hypereconomy

[Version 0.46 - 3-March-1998]

© 1997 Alexander Chislenko -
[This is my original text on Hypereconomy. Please see Hypereconomy development group page for other materials on this topic]

Multi-agent systems and action spaces

People distinguish various types of systems consisting of relatively autonomous functional parts. They call them: ecology, economy, society, organism, software, etc. The rules for identifying these systems are quite simple. For example, any system consisting of biological objects would be called an ecology, one made up of humans is a society, and a part of the latter that has to do with exchanging things is an economy. However, if one wants to understand the behavior of such systems (otherwise, why study them?) it seems more productive to subdivide them into functional types. This also allows to draw analogies between similar systems of different nature (e.g., a communist economy may be similar to an integrated organism, as its parts work directly on the central order, and the role of financial flows is reduced to assist the centralized regulatory mechanism. A contract economy looks much like a forest ecology, except that agents get locked into relationships by contracts, not niches, and some other manifestations of subsystems' intelligence. So multi-agent systems can be classified by overall complexity, internal control methods, structure and viscosity (transport capacities) of the space, methods of connecting agents into stable working groups (mutual specialization, physical coupling, force, emotional affection, ownership, contracts), diversity and intelligence of agents, etc.

A general theory may also help to predict where new systems may have features qualitatively different from everything previously observed. For instance, space within modern communication and computing environments is effectively non-metric (it takes approximately the same time and effort to phone anybody in a country, or address any memory unit within a computer). This means that most space-related laws of all previous functional spaces would not apply to "digital" systems: there is no concept of premium location here, or differential rent, no "reproductive isolation" or other influence of space-induced separation on agents' communication and diversity. Also digital systems often have lower replication costs of agents than execution costs, which makes them dramatically different from all systems more essentially embedded in their physical substrates. - For humans, it's much easier to translate an article or to fix an appliance than to transfer the necessary skills to another person. For a computer it is just copying a program.

Together, the above features create a foundation for system design that is dramatically different from all natural and human agglomerations, in terms of agent specialization, diversity, and deployment. I would envision the future intelligence as a collection of highly specialized tools, with high-bandwidth interconnections, with "personalities", or complex problem solvers ("achievers", "goal engines", "research tools", "development threads") consisting of limited contractual relations between multiple problem-solving, planning, perceiving, and acting entities, that can at the same time be employed in many other similar relations. So the entities may use the same "body parts" on a cooperative or time-sharing basis.

Economy and beyond

Economic relations are based on a new class of signaling instruments. By using quantitative symbols to represent typical value relations between various resources, economy allows efficient distribution of signals in the system, and provides short-sighted selfish agents with incentives to adjust their consumption and production efforts in accordance with the collective requirements of all other agents in the system.

While the economy provides its agents with accurate estimates of generic values of typical goods and services, it has little to offer to each individual agent in their attempts to estimate the utility of a certain product for their particular needs, or offer any other personalized or situational advice. This kind of knowledge is usually obtained by the economic agents in about the same way they discovered average social values of common products and services in pre-economic times: personal experience with the environment and direct communication with other agents. There are no quantitative instruments available for automated processing for this purpose.

This is understandable: comprehensive detailed data about each agent's behavior so far has not been available in a usable form, communication and computational tools necessary for processing this data have not been adequate, transaction costs have been too high, etc. All these obstacles are now being removed in many systems.

With these tools in place, there appears a foundation for "a second signaling mechanism" in an economic system: together with conventional aggregate indicators of average costs of resources and services and their expectations, the new crop of signaling instruments may deliver suggestions of expected value that particular agents may derive in their individual situations from such resources or services. Then agents can be better equipped to optimize their behavior by using all available knowledge for comparing personal and global costs, risks, and utilities.

Systems carrying representations of situational knowledge ("hyper-economies", or "super-economies") may be implemented in both completely automated environments and those involving humans (e.g., they could assist people in selection of things most appropriate for their tastes and situations, from recommending movies they want to see, to tools that would satisfy their needs at the best price to choosing diets and insurance plans for their own goals and conditions).

There is some analogy here with the notion of "psychohistory" suggested by Isaac Asimov in his Foundation series, except that instead of using detailed representations of social processes for long-term prediction of future events (which is hardly possible with a still very imperfect model of a huge chaotic system), I suggest a more modest goal of using them for richer local modelling and adaptive control. The only advantage here is that this modest goal seems actually achievable.

From research to development

I know of a very few people working in similar directions. In the current situation, good theorists work on conceptualizing existing systems, and are usually not aware of new developments in software that may create foundation for new systems and theories. On the other hand, people working on the first versions of new signaling systems usually do not have much expertise or interest in theoretical analysis of complex systems, and concentrate on debugging and distributing software. As a result, the new system structures are arising, once again, unintentionally.

I am trying to approach this topic both theoretically and experimentally.

Theoretical part consists of considerations of how the new signaling systems may change our familiar economic notions, such as capital, interest, efficiency, liquidity, inflation, emission rights, secondary and derivative instruments. Some of these features will remain relatively unchanged but become more liquid (faster distribution and lower transaction costs) and more efficient. Other parameters may turn from scalar to vector or a more complex structure (e.g., augmentation of the numerical representation of value of an object with a matrix of its utilities for various purposes as estimated by different agents. Some features, such as non-scalar "situational/detail derivatives" for assessing values and risks of, and balancing among, particular utilities will be entirely new.

Similar changes will happen with the concepts from social and ecological fields. Communities do not have to be local anymore, they may have variable/adaptable geometries, relations between managing and performing agencies will be redefined, new control structures and dynamic patterns arise, etc.

On the experimental side, we are working on simple versions of such systems in Internet environments, where the data is already available, and the results seem of direct practical use. Recently, we created a Hyper-Economy Development Group (HEDG) for modelling such systems. The group is open to interested researchers.

Among the first applications of this approach I envision large automated collaborative filtering projects, link exchange programs and targeted advertising, balancing flows of information in accordance with diverse and dynamic interests of creators, consumers, providers, and sponsors of the Web content.

My hope is that eventually such hyper-economic schemes will develop to store and process increasingly complex representations of agent interests and utilities, and have a great variety of competing intelligent algorithms for optimizing system processes, thus turning into a hybrid of an economic self-regulatory mechanism, complex multifaceted community, and distributed artificial intelligence, with secondary data representation and signaling mechanisms forming emerging self-awareness of the new systems.

It seems noteworthy that in these constructs, as well as in conventional economy, the meta-signaling [generalized/abstract in relation to actual "field agents"] level is represented in terms of signal currents in the semantic space rather than crisp symbolic constructs. We can interpret this as a sign that the connectionist approach is natural for a fetal stage of any system intelligence. Or - better - that a combination of symbolic processing with signal flow patterns will extend the human balance of symbolic reasoning and intuition to further levels of intelligence.


A few examples on how one may combine general, specialized (or situation-specific) knowledge: (actually, there is hardly any absolutely general knowledge that is equally applicable to all situations; in many cases, situational advice may consist of references to relevant general knowledge).
Recommendation systems for consumer items:

The price of an item, such as a book, an instrument, or a travel package, may be taken from the general catalog. The estimates of utility of this items would come from the collective experience of other people with similar needs and interests who have experience with this or similar items. Some computational constructs needed for summarizing and delivering situational knowledge, such as utility trends on item/consumer clusters, vectors of refined "eigen-preferences", etc., represent hyper-economic equivalents of secondary and derivative signaling instruments in the conventional economy.

Targeted advertising:

The same mechanisms that can advise items to consumers, may suggest potential consumers to manufacturers of these items. In the current situation, an advertiser knows the exact price of the ad delivery, but targeting of ads is usually very costly and inaccurate. As a result, most of the multi-billion-dollar advertising industry efforts miss their targets and annoy people, while knowledge of many good products doesn't reach the individuals who may benefit from them. An advanced mechanism for combining and sharing (for a price) metabolites of advertising and sales information of all companies would result in more efficient proliferation of useful goods and services and dramatic reduction of targeting costs.

Computational heuristics:

In a large AI system, computational/royalty costs for all algorithms may be known, but it is also important to estimate the expected benefits of applying each algorithm to each problem. In an "intelligent soup" of a "market AI" system, multiple problem generators and solvers can share not just information on the current value and availability of computational resources, but also suggestions of what methods appear to be of the highest utility for what problems, and how they may be better applied.



The latest hypertext version of this article is available at