Figure 1: The Structure of the IBM Information Filtering Model
Scott Moss*
Centre for Policy Modelling
Manchester Metropolitan University
http://www.cpm.mmu.ac.uk/~scott
s.moss@mmu.ac.uk
2 How economics sets a bad example.
3 Some consequences of adopting agent representations from economics
4 Intermediation in information filtering markets.
5 An extended information filtering economy model
*This paper arose from a series of discussions
with Ed Durfee, Aaron Armstrong, Moshe Tannenholz, Michael Schroeder and
Leigh Tesfatsion, none of whom necessarily agree with the argument or my
conclusions. I am grateful as always to my colleague Bruce
Edmonds with whom everything in the paper has been discussed. The
intellectual imporatnce of SDML will be apparent from the paper.
SDML has been designed and developed by Steve Wallis in the VisualWorks
development environment. It also runs on top of the VisualWorks 2.5.x
object engine. The development environment and permission to distribute
the object engine free of charge for academic purposes are provided by
our sponsors, ObjectShare (UK) Ltd.
The purpose of this paper is to demonstrate that the effort is misdirected and that to hold to long-standing, demonstrably effective principles of software engineering better enables the power of the agent paradigm to be realised.
The factual position regarding the nature of economics and the reasons why economics sets a bad example for computer science are explained in the section 2. An example of the consequences of adopting economists' methods is presented in section 3. A constructive alternative approach is offered in section 4 with a demonstrator implementation described in section 5. The results from that implementation are reported in section 6. The wider lessons to be drawn for multiagent systems research more generally are argued in section 7.
A seminal work in this vein was Thomas Sargent's Bounded Rationality and Macroeconomics [10]. Sargent addressed an implausibility in a core concept in conventional economics called rational expectations. A rational expectations equilibrium is one in which every agent in the system knows the correct model of the system. The common application is to macroeconomics so that everyone in the economy knows the forecasting model that will correctly predict the future values of all macroeconomic variables. It is usually assumed in addition that every agent has the same utility function – this being the "representative agent" assumption that diversity of preferences or expectations within the economy have no significant or biasing effect on its performance in aggregate. This assumption is wholly unsupported by any evidence whatsoever. On the basis of Sargent’s, there is a growing literature in which agents are represented by some implementation of a genetic algorithm and the goodness of the models incorporating this representation is measured by the closeness of its output to a rational expectations equilibrium.
The important point here is that both the agent specifications and the measure of goodness of the models are not at any point considered in relation to any observations of the real world.
The consideration of markets in modern economics has a much longer history. In the 1880s, Leon Walras devised the general equilibrium concept in which there is a set of prices such that the demands for and supplies of every good and service produced and traded in the economy were equal. There was no excess supply of any good or service and no excess demand. The process by which this would generally be achieved has never been identified. Walras conjectured that such an equilibrium could be found by a process of tâtonnement in which an auctioneer would receive all bids for and offers of goods and raise the prices of those goods for which the bids exceeded the offers and reduce the prices of those goods for which the offers exceeded the bids. This process is unstable whenever there are more than two goods and two traders in the economy. A recent attempt by Binmore et al. [2] to devise a game theoretic model was claimed to be interesting because software agents face similar problems when they meet other software agents that are "badly programmed" meaning that they are not rational in the sense that economic agents are assumed to be rational. The agents in that game were to be considered "as a metaphor for an evolutionary process in much the same way that the auctioneer of neo-classical economics is a metaphor for some unmodeled process that eventually equates supply and demand."
This is the crux of the issue. Whether in game theory or in market models, the agents and the processes specified in economic models are metaphors for some unjustified impression invented by the modellers out of their own heads. These arbitrarily specified agents are engaged in arbitrarily specified processes in the precise sense that both agent and process specifications are entirely unvalidated. What is more, verification takes the form of nothing more than ensuring that sufficient (but not necessary) conditions exist for an equilibrium that itself has no empirical correspondence to any observations whatever. Economic theory is a collection of metaphors for unmodelled (because unobserved) processes and unobserved agents.
While these metaphors have been translated into a number of areas of multiagent systems research, the potential consequences of such translations are readily demonstrated by one example. The example developed in this paper is that of the information filtering economy in which automated information agents buy and sell information ultimately to provide human users with sets of information tailored to their individual interests and requirements. The inspiration for these information markets is the partial equilibrium theory of microeconomics.
In this theory, competitive markets are defined on a set of agents who demand a good or service and mutually exclusive set of agents who supply the good or service. All agents on both the demand side of the market and the supply side of the market are price-takers. That is, they observe the market price which is the independent variable of their demand and supply functions, respectively. If there is no production, then the demand and supply functions are derived from utility functions. The market is cleared (demands and supplies are brought into equality) by effectively instantaneous price adjustments such that, in the stable cases, a price increase reduces excess demands and a price reduces excess supplies.
Of course, we do not observe many such markets and those we do observe are the organised stock and commodity exchanges that invariably have very strict rules of behaviour and stylised mechanisms for the exchange of assets. There are in these markets both jobbers and brokers where the jobbers buy and sell the assets at prices they specify while brokers arrange transactions between sellers and jobbers and between buyers and jobbers but do not themselves buy and sell. Other markets patently do not work that way. Commodities that are specialised and produced for thin markets (such as ships) are only produced to order and are never produced to hold as inventories. The price of each unit produced is negotiated in advance. The same is true for specialised buildings such as sports stadia but less true for buildings that can be well standardised and for which markets are densely populated such as houses on housing estates which might differ only by colour. Chocolate bars, domestic soaps and other fast-moving consumer goods are sold by producers to wholesalers who sell them to retailers who sell them to the users. Large retailers can bypass the wholesalers.
Evidently, common observation indicates that the processes of exchange for different commodities are simply different.
There are three specific questions to be raised in this regard:
The objective of the IBM model is to inform a process whereby "agents will progress naturally from being mere facilitators of electronic commerce to being financial decision-makers, at first directly controlled by humans and later with increasing autonomy and responsibility". The rationale for adopting the perspective of economics is that
A schematic description of the IBM model is given in Figure 1.
Figure 1: The Structure of the IBM Information Filtering Model
In this model,
The authors point out four restrictions imposed on the model that, presumably, are to be relaxed to capture usefully scaled implementations of information filtering economies. These are
In all of these cases, as in economic theory, agents have to know the actual state of the world (or market) and the actions (or even more strongly, the utility functions) of every other agent. In order to relax these assumptions, the IBM team employed simulation techniques that indicated the system outcomes to be sensitive to the assumed initial conditions.
In summary, the IBM model of an information filtering economy entails known restrictions in order to render it tractable while adding an element of realism renders the model both analytically intractable and sensitive to initial conditions. This is a consequence not only of the rational agency assumption but the model-building procedures required by that assumption.
In building models with rational agents, there are inherently restrictions on the specification of the systems in which they are implemented. While the specification of any agent will entail some restrictions on the systems, the restrictions imposed by rational agency as evidenced in the IBM information filtering economy are undesirably (and unnecessarily) strong. An alternative is to start with the specification of the desired system and implement agents that can function usefully in that system. In the case of markets, a weak specification of agents has been known for twenty years [7, 8] to support some very general and robust analytical results on the institutional structure of markets. These results are exhibited and applied to an information filtering market in the next section.
The first issue raised in this regard concerns the conditions in which any intermediary can function in a market. In order to get a handle on this question, the weakest possible assumptions are offered about agents. It is assumed that
A2 agents prefer survival in the market to exit from the market
Taken together with elementary accounting identities on cash flows, assumptions A1 and A2 have the following consequents:
C2 agents will always seek to increase their sales revenue if to do so involves no additional expenditure
These two assumptions and the two consequents imply the fundamental law of market intermediation:
 |
The price at which an intermediary offers an item (the offer price) |
 |
The price for which an intermediary bids for an item (the bid price |
 |
The price in direct exchange (paid by a user to a producer without the intervention of an intermediary) |
 |
The transactions costs incurred by a user in purchasing an item in intermediated exchange |
 |
The transactions costs incurred by a user in purchasing an item in direct exchange |
 |
The transactions costs incurred by a source in selling an item in intermediated exchange |
 |
The transactions costs incurred by a producer in selling an item in direct exchange |
 |
The transactions costs incurred by an intermediary in selling an item |
 |
The transactions costs incurred by an intermediary in buying an item |
It follows from consequent C1 that an agent will purchase from an intermediary only if the total cost (offer price and transactions costs) of buying from the intermediary are less than the price and transactions costs in direct exchange with a source. That is,
(1) 
From C2, an agent will sell to an intermediary only if the price net of transactions costs is greater than in direct exchange. This implies
(2) 
Combining these inequalities,
(3) 
Applying A1 and A2 to the intermediaries and solving in expression (3) for the intermediary’s bid-offer price spread,
(4) 
which states in effect that any intermediary’s transactions costs must be less than the bid-offer spread which must be less a fortiori than the total savings in transactions costs by sources and users as a result of the activities of the intermediary. Consequently, total transactions costs are smaller if intermediaries can survive in a market than if they are absent. Since the bid-offer spread can only appropriate a part of those savings, the total costs of trading in any market will be reduced by the profitable presence of intermediaries.
This reasoning leads to the following hypothesis on intermediary survival:
The extension required is one that does not suffer the restrictions imposed and acknowledged by the IBM team. That is, it will have the following characteristics:
Figure 2: The Structure of the Extended Information Filtering Model
The roles of the agents in this model are:
Exchange cannot take place without communication between buyer and seller. The model was implemented in SDML, the strictly declarative language that supports representations of communication by agents to assert clauses to one another’s’ databases. It also supports nested time levels and, within time levels, agents can act in parallel. In order not to change the state of the world for one agent by virtue of the actions of another agent, no clause asserted by one agent will take effect for any other agent until the next time step. Consequently, if agent-i asserts the clause (demandEnquiryFrom agent-i [3 7 2 6]) to the database of agent-j, then agent-j will not be able to access that clause until the next time step.
All cognitive agents in the model are synchronous parallel agents. To enable them to communicate in order to seek, agree and effect transactions, within each major time step, the trading cycle, is a nested time step, the communication cycle. A limit of six communication cycles were allowed per trading cycle though there were fewer communication cycles if all demands were filled earlier.
Each source was allocated at random a number and selection of positions on the information string, In each trading cycle it stored a pairlist of indices of the string and the current value at the indexed position. The digit string length, its number base, the number of sources and the maximum number of items of information that could be held by each source are specified by the model operator for each run.
The endorsements schema are the key distinction between the agents implemented in the system reported here and rational agents. The endorsements schema are an alternative to utility functions. The principal conceptual difference is that an endorsements scheme is a framework for agents to reason and learn about preferences for different objects whereas utility functions represent static and unchangeable preferences by agents.
An endorsements scheme links mnemonic tokens to numerical values with a basis of comparison among different values. Clauses defining endorsements schema of information users in the extended information filtering economy model include:
endorsementSchemeDefinition infoSourceEndorsementScheme
[[lowCost 1] [highCost -1] [mostCompleteInfo 2] [accurateInfo 3] [inaccurateInfo
-3]] 1.2\
(5) 
where the ei are the values of the endorsements on the object and b is the third argument of the endorsement scheme definition clause.
An example of a rule endorsing an object is reproduced in Figure 3 from the main rulebase of type InformationAgent. The antecedents are satisfied if it is the first communication cycle of the trading cycle and if there was a purchase of information items from a supplier during the previous trading , items (unified with ?content) were supplied during the same trading cycle and there were no items purchased but not actually supplied. If all of that is true, then the clause endorsing that supplier as reliable is asserted to the database corresponding to the time level tradingCycle so that it does not have to be asserted each communication cycle.
| and
time tradingCycle ?tc\ last tradingCycle (and purchaseFrom ?supplier ?indices ?price\ infoSuppliedBy ?content ?agent)\ not (and includes ?indices ?index\ not includes ?content [?index ?value])\ |
|
|
Both utility functions and endorsement mechanisms relate to agent preferences. The difference is that a utility function specifies a map of possible outcomes and preference relations among them while the endorsements mechanism is a process of developing preferred courses of action based on and changed by experience. It is not surprising that to impose fixed preference maps on agents and an algorithm to be used for making decisions whatever the prevailing circumstances, will restrict the domain of application of the system to a class of states in which the algorithm is applicable or tractable and exclude all other states. By defining a process that is itself contingent upon the state of the system, the endorsements mechanism can respond to emergent system phenomena as demonstrated in the extended information filtering economy model.
The endorsements mechanism generates emergent preferences among suppliers based on the behaviour of the suppliers and the nature, for example volatility, of the information they offer.
More direct means were used to ensure that agents did not buy from high-priced brokers. In the initial trading cycle there were no brokers and information users continued to search among information sources until all available information items in their lists of demands had been acquired. The sum of the transportation and processing costs of those acquisitions were retained in permanent memory. If, in any subsequent trading cycle, an agent received an offer of supply from a broker at a price which made the items wanted more expensive than the cost at which they were previously obtained from an information source, then the broker was endorsed as being a high-cost supplier and the current value of the information would be obtained from the source instead.
At the same time, brokers would seek to obtain high prices by building models of the effects of their mark-up strategies. The mark-up for each broker was applied to the cost of supplying the information items it expected to sell on the basis of demand enquiries in hand. When demand was strong, the mark-up would typically rise and when demand was weak it would fall. As will be seen from the simulation results, even the most successful brokers would occasionally see their prices undercut by a new, strategically placed broker so that they lost all custom and rarely recovered. In setting prices, each broker would note the maximum price the customer claimed it would pay and set an actual price per item as a random linear combination of the mark-up on anticipated costs and the maximum price.
6. Simulation results: testing the intermediary survival hypothesis
There are no production costs in the model as specified for this paper. Consequently, all costs are transactions costs resulting from transportation and processing as modified from the IBM model.
The test of the hypothesis on intermediary survival is whether intermediaries survive when their presence reduces total information costs and not otherwise. From the analysis of section 4, intermediaries can survive in the market only if there are economies that they can enjoy that cannot be enjoyed by either sources or users. The special advantage of brokers in the extended information filtering economy model is that they can break bulk (which the sources cannot) and they can sell information (which the users cannot). Consequently, brokers will be able to reduce the transactions costs of the users by selling them the information they want and not requiring them to process information to determine whether they want it. Also, brokers will be able to reduce transportation costs by locating close to the users and incurring the bulk of the transportation costs once and then lesser transportation costs on to each of the users rather than multiplying all of the transportation costs from close to the sources.
One would anticipate that location would be relatively important when transportation costs dominate transactions costs and that location would be relatively unimportant when processing costs dominate.
In the original publications of the markets analysis described in section 4, Moss [8, 9] argued that intermediation would also require the number of intermediaries to be small in relation to each side of the market. Common observation indicates that large chains of supermarkets and department stores, for example, purchase directly from producers while smaller, independent shops purchase from wholesalers. If this argument is correct, reducing the number of information users or the number of information sources in the extended information filtering economy model would reduce the scope for intermediation by reducing the scope for savings in information costs.
The parameters selected for the various simulation runs were set to test these hypotheses. Because the information sources are not cognitive agents, there was no point in varying their properties. The main aspects of comparison involved different transportation and processing costs, different rates of entry by brokers into the market and different numbers of information users.
It was common to all simulation runs that the information digit string had 40 digits to the toroidal number base 8. There were 15 information sources each of which could hold a maximum of 15 information values. Each information user could demand up to 12 items of information each trading cycle. The maximum number m of new jobbers that could enter the market was either 1 or 4 or 8 so that at each trading cycle the number of entrants was drawn at random from the U[1,m) interval. The transportation costs of one packet over the unit distance was either 1 or 1000 and the cost of processing an information item was either 100 (if the unit transportation cost was 1) or 1 (if the unit transportation cost was 1000) or, to test the effect of increasing costs 1000. The number of information users was either 100 or 25.
In order to determine the effects of intermediation in the market, no brokers were active during the initial trading cycle of each simulation and the cycle only ended when every information user had acquired the available information it demanded.
Figure 4: Information cost indices corresponding
to
different degrees of ease of entry by brokers – 100 users.
The first issue addressed was the effect of ease of entry on the efficacy and cost of information filtration. As seen from Figure 4, in the case where the unit transportation cost was 1000 and unit processing cost was 1, there were no significant differences in information costs among the three runs with maximum entry at 1, 4 or 8. The vertical axis measures the information cost index such that the index = 100 at the initial trading cycle. Evidently, the system had settled down by trading cycle 9 after which information costs oscillated mainly between 50 and 60 percent of the costs in the absence of intermediation. There is no obvious reason to conclude that increased ease of entry, as represented by higher average numbers of entrants, reduces information costs. After the first 10 cycles, the average information cost indices for maxima of 1, 4 and 8 entrants, respectively, was 54.74498, 56.93128 and 53.40589 with standard deviations of 6.537899, 6.63393 and 9.051949. The differences in the means are neither ordered according to ease of entry nor statistically significant. There is a possible relationship between variance and ease of entry since the standard deviations increase with the maximum number of entrants. There are too few observations to reach a conclusion on this conjecture.
Figure 5: Total information cost index: 25 users
The markets analysis suggests that a smaller number of information users will reduce the scope for economies of scale available only to intermediaries so that there will be fewer intermediaries and less cost savings. A run with 25 information users and a maximum entry of eight brokers per trading cycle generated the series for the total information cost index in Figure 5.
It is clear from Figure 5 that total information costs are systematically higher than in the initial trading cycle when there were no intermediaries. Total information costs are higher they include acquisition of information directly from sources by both brokers and users. That no intermediary survived its first trading cycle indicates that all failed to recover the costs of their own information acquisitions. The loss-making information sales to users nonetheless accounted for a substantial proportion of user purchases. Indeed, by this measure of market penetration, there were no significant differences between the run with 25 users and any of the runs with 100 users. The data series on market penetration by brokers are reported in Figure 6.
After trading cycle 5, no broker survived a single trading cycle. Of the six brokers that survived beyond the time of entry in the first five cycles, two survived into a second trading cycle, two into a third and one each into a fourth and fifth. None of these survived past trading cycle 8.
By contrast, intermediaries did survive in runs with larger numbers of users where the information cost index averaged little more than half the value without intermediation. These results are reported in Figure 7 and Figure 8.
With entry by one broker per trading cycle, the entrants at trading cycles 1, 3 and 4 survived profitably through the whole simulation run of 50 cycles while no other broker survived at all. The effect of increasing the maximum number of entrants per trading cycle is seen in Figure 7 and Figure 8 where the maximum entry was 8. These figures report the market shares by volume of jobbers surviving at least into a second trading cycle with at least a five percent market share in at least one trading cycle. Figure 7 is derived from a simulation run with high transportation costs and from a simulation with high processing costs. The "high" processing costs were set at 100 per item received because the average number of items in a packet was on the order of 10 (maximum 15). In consequence, the "high " processing cost per information packet was about the same magnitude as the "high" transportation cost per information packet. Evidently, there will be a small number of dominant brokers in the market specified here. Under some cost conditions there is a trend towards monopoly though the monopoly need not be permanent.
Figure 7: Market shares by volume of brokers in
high transportation cost regime
In the run generating the data for Figure 7, the first monopolist, exchangeAgent-1-2, lost its monopoly at trading cycle 17 to the new broker, exchangeAgent-17-3, and after a brief struggle was left with only sporadic sales on a small scale. Its subsequent survival was due entirely to the substantial (but declining) financial resources from its erstwhile monopoly. Towards the end of the simulation run, competition was increasing as exchangeAgent-36-1 and exchangeAgent-40-2 established footholds in the market. The randomly generated location of each of these brokers was close to the information users’ edge of the market space and to each other. The mark-up by exchangeAgent-17-3 on costs to determine its minimum prices was less at entry than the mark-up of exchangeAgent-1-2. And the mark-up of exchangeAgent-40-2 at its entry was less than the prevailing mark-up of exchangeAgent-17-3. The proximity of these agents to one another meant that their costs of supplying users were much the same. The smaller mark-up of the entrant meant that its prices were lower and this was sufficient for it to establish a strong position with the customers of the incumbent.
Figure 8: Market shares by volume of brokers in
high processing cost regime
The results reported here unambiguously confirm the basic hypotheses: intermediaries can survive if and only if their activities reduce total costs in the market. In the information filtering market modelled here, all such costs are information costs. The original markets theory published some 20 years ago indicated that larger numbers of users would enhance the scope for intermediaries to reduce transactions costs and this result, too, has been confirmed. However issues concentrated on by economists, such as ease of entry into a market, appear to influence the patterns of market shares but not to influence the effects of intermediation on costs. A natural conjecture to offer in this regard is that limited cognitive capacities of information brokers will generate some relationship between numbers of surviving brokers and the scale of the information system as measured by numbers of information sources and users. This is not, however, an issue related to competitiveness. The claimed benefits of ease of entry for competitiveness and efficiency do not seem to apply to market intermediation.
The model and results reported in this paper were developed from the specification of a desired software system with a specification of information agents that was compatible with the needs of that system. Properties of the system were derived from first principles and relevant, common observation of the real-world phenomena that inspired the software system by analogy.
The endorsements mechanism of the agents supported the development of preferences by experience in a manner that was sufficiently flexible to enable them to function under a range of cost and competitive conditions. By contrast, the specification of "rational" agents in the IBM model is the likely source of the need undesirably to restrict the specification of that model.
There remains the question of how robust are the results reported here with respect to different agent specifications affect those results? While the answer is obviously beyond the purview of this paper, one approach to the investigation of that issue follows naturally from the use of the SDML modelling environment [9].
SDML is a strictly declarative language that corresponds to a fragment of Konolige’s strongly grounded autoepistemic logic. This correspondence is the result of a co-evolution of the language and modelling approaches developed by the Manchester Centre for Policy Modelling. It has proved possible in simple cases to use the language to prove theorems about models’ properties. More powerful theorem provers would doubtless extend the ability to prove such theorems.
Any model that runs under any programming language is sound and consistent with respect to that language. Models developed in SDML, including the model reported here, are sound and consistent with respect to SDML and therefore to strongly grounded autoepistemic logic. Consequently, the models (though not necessarily the agents) have the same formal clarity as agents represented by procedural formalisms such as expected utility maximisation or declarative formalisms such as BDI or deontic logics. Since the purpose of developing multiagent systems is to generate useful, reliable systems using agents as a powerful programming analogy, it would seem more appropriate to ensure that the properties of the systems rather than the agent representations are robust. For this reason, it would be useful to prove theorems about the effects of, for example, different degrees of ease of entry, different numbers of users etc. in order to determine wherever possible what properties of the system are independent of the properties of particular agent specifications.
These are, of course, issues for future research. For the present it is sufficient to have demonstrated that the cost of rational agency is a loss of scalability and applicability by virtue of the requirement to specify systems that such agents can support.
[1] Binmore (1987), "Modeling Rational Players – 1", Economics and Philosophy, Vol.3, pp.179-214
[2] Binmore, K. M. Piccione and L. Samuelson (1997), "Bargaining Between Automata" in Rosaria Conte, Rainer Hegselmann and Pietro Terna (eds.), Simulating Social Phenomena (Berlin: Springer-Verlag, Lecture Notes in Economics and Mathematical System), pp. 113-132
[3] Cohen, Paul R. (1985), Heuristic Reasoning: An Artificial Intelligence Approach (Boston: Pitman Advanced Publishing Program).
[4] Hanson, James E. and Jeffrey O. Kephart (1998), "Spontaneous Specialization in a Free-Market Economy of Agents"
[5] Kephart, Jeffrey O., James E. Hanson, David W. Levine, Benjamin N. Grosof, Jakka Sairamesh, Richard B. Segal and Steve R. White 91998), "Dynamics of an Information-Filtering Economy" in Proc. Second International Workshop on Cooperative Information Agents (CIA-98).
[6] Konolige, K. (1988), "On the relation between default and autoepistemic logic", Artificial Intelligence, v. 35, pp. 343-382.
[7] Moss, Scott (1981a), "Technology and Market Structure in the Analysis of Exchange", Managerial and Decision Economics, v. 2, pp. 106-120.
[8] Moss, Scott (1981b), An Economic Theory of Business Strategy (Oxford: Basil Blackwell).
[9] Moss, Scott, Helen Gaylard, Steve Wallis and Bruce Edmonds (1998), "SDML: A Multi-Agent Language for Organizational Modelling", Computational and Mathematical Organization Theory v. 4, pp. 43-70.
[10] Sargent, Thomas (1993), Bounded Rationality and Macroeconomics (Oxford: Clarendon Press).