The Fishmarket Project


In this project we are concerned with communicational aspects of multiagent systems. Two issues center our attention: the dialogical aspects of communication among intelligent agents, and the accountability of agent interactions.

We advance some ideas on how a layered agent model can be embedded in a multi-agent system, and how such a system can be grounded on basic communicational conventions and a shared ontology through what we call a dialogical framework. We also propose a characterization of an agent-mediated institution as a set of linguistic, ontological and behavioural constraints that allow heterogeneous agents to interact.

And in order to explore and test these ideas we develop network-based market place prototypes based on the trading conventions present in the traditional fish market which is regarded as a multi-agent negotiation environment in which to test agent models.

What is the (real) fish market like Picture the situation:

The village fleet is unloading. The sea-weathered fishermen's faces are starkly carved by the slanting afternoon light. Shining slivers of living silver are slipped into coarse wooden boxes. And while buyers inspect the day's catch with carefully studied displiscense, boats, gears, nets and bystanders occupy their proper places in what, by all evidence, seems a well-rehearsed choreography.

Suddenly, a loud, clear voice resounds under the high vaulted ceiling. It is the master auctioneer calling for attention. The anticipated litany starts: Cuatro arrobas de sardinete: 300, 290, 280,.... A quick gesture by the woman next to you brings the series to a halt; a discrete ruffle propagates in all directions and subtle quick gestures indicate some sort of public acknowledgement of a transaction. The master bidder re-enacts the call once more, and once more, and over and over again the by now familiar process of interrupted arithmetical series is replayed until the last box of fish is sold.

You are witnessing the time-honoured institution of La Lonja del Pescado, the downward-bidding fish market of Spain. A typical commodities-trading marketplace. A social convention of undeniable utility and misleading simplicity.

At first sight, the fish market (or its many variants) is a place where goods are exchanged. But a closer look reveals also a socially adopted ritual in which individuals perform well-defined roles. One main action, trading, is performed by a collectivity of participating agents. A form of negotiation takes place within that crucial action: a box of fish (which originally belonged to a fisherman or to a group of fishermen) is sold to a buyer at a publicly agreed and acknowledged price. But the negotiation is actually performed under a well defined protocol in which it is not the fisherman who offers the fish, but an intermediary, the auctioneer, who calls for tenders and who adjudicates the purchase to the highest bid --in fact, to the buyer who first calls a price.

The concatenation of many rounds of such negotiations constitutes a day's work at the fish market, but the ritual involves other subsidiary actions as well. Take the auctioneer's role, for instance. She not only calls the goods and states the sequence of prices, among other actions she actually presents the goods to be sold and sets the starting price, and by so doing establishes bounds to the negotiation; she recognizes the actual buyer and the actual buying price, and she also uses the last bidding-round information as a guideline for the selection of goods and prices for the next round. Buyers are involved also in other actions, certainly they express commitments to purchase boxes of fish, but they also (for instance) participate in a process of accreditation whereby they are admitted to the Lonja as bona fide buyers, only if they can prove solvency either by being supported by already admitted buyers, or by posting a bond or some other form of guarantee.

Institutions, Dialogues and Shielded Agents

Arguments in favor of building an electronic auction house

Internet is likely to spawn many new markets. One that is particularly attractive for multiagent technologies is network-based trading .

But if that potentially important market is to become a profitable actual market two issues need to be addressed. First the issue of diversity: how to contend with the diversity of goods, trading conventions, market requirements, timing, etc. that are likely to be present in network based markets. Second, the issue of performance: how can trading and negotiation be accomplished in an effective, practical way; what roles can and should software agents play (what level of rationality is required of those agents and how can it best be achieved in order to aptly play those roles); and, finally, what kind of accountability needs to be required from trading agents and how is it going to be assessed.

Instead of developing specific solutions to the problematic aspects of those issues, we propose to circumvent many of them by imposing simplifying constraints on participants' interactions, and building electronic intitutions in which these interactions can properly be accounted for.

An agent-mediated institution is an electronic environment where certain kinds of goods are traded under explicit trading conventions. These environments are restricted in the sense that the trading conventions impose explicit objective requirements to the admission of goods and traders into the market-place, and constrain the agents behaviour --within the market-place-- to only those actions that are necessary and sufficient for actual trading under those conditions. Essentially the same purpose that is served through the socially adopted conventions of the traditional auctions and other institutional markets.

A second --more technical-- motivation:

Dialogues are pervasive and, to a large degree, unavoidable. Legal arguments, political debate, domestic disputes, didactic explanations, interviewing, psychotherapy, coordination of actions, negotiation, all tend to involve some form of dialogical interaction. But dialogues are also unavoidable since what is accomplished through them cannot be accomplished in a strictly monological setting, because some fundamental ontological, rethorical or epistemic features would be lost. Thus, dialogues may be worth studying, although they are not simple entities. Certainly not from a formal perspective.

On one hand, dialogues involve multiple participants, who exchange illocutions in rich and complex languages. Thus, classical --i.e. monological, truth-semantical, non-dynamic-- formal devices, are inadequate to deal with these complexities. But in addition, dialogues are typically situated (or opaque or unstructureded), in the sense that participants react to the illocutions, depending on the conditions or elements present in a given context or situation. In many dialogical situations meaning is not necessarily established in an objective,a priori, form; nor are interventions subject to an objective, a priori, clearly expressible protocol. In typical dialogues, participants confirm, adjust, refine or establish their own meanings, intentions, beliefs and actions according to their individual interpretation and what the other participants are saying. The first kind of complexity has been addressed through ad-hoc dialogical structures [ c.f. Hamblin, Rescher, Hintikka, Carlson for different approaches], the second one has been the object of increasing attention, mostly from the idea of a situation [c.f. McCarthy & Hayes; Reiter; Georgreff; Sandewall; Barwise & Perry; Barwise; Devlin], but also from the notions of speech acts and --closer to our concerns-- conversation [Wittgenstein; Searle; Winograd & Flores; Singh; Haddadi]

Certainly there are dialogical situations which are irremisbly opaque. That is the case, for instance, of psychotherapy or everyday conversation, where meaning and commitments are mostly established through highly unstructured dialogical interactions. But then there are other contexts --such as auctions and other similar forms of mediated trading and structured negotiation-- where a priori univocal shared interpretations (transparency) is not only desired but enforced .

From a formal perspective, structured negotiation environments still contain intensional, structural and functional elements that are characteristic of complex dialogical contexts, but lend themselves to a much more straight-forward treatment because of their explicitness.

Formal tools for dialogical frameworks

In order to give a formal account of agent-mediated interactions, we will construct a syntactic dialogical framework and introduce the corresponding semantic contextual interpretation. These constructs will attempt to formalize the intuitive notion of a transparent context, as a way of interpretting agent illocutions within a given environment. We will use them also to express in objective terms how illocutions are exchanged among agents, and what effects these exchanges have with respect to that context.

Several ontological commitments need to be made explicit in order to handle illocutions . For instance, different classes of agents, agents' beliefs and commitments-to-act, time of utterance of the illocution, etc. Consequently, not only a rich and expressive enough language has to be available, but also appropriate formal devices need to be developed to represent the corresponding intensional contents of the illocutions --truth, belief, action. Furthermore, in structured interactions the interweaving of illocutions follow definite patterns. Thus, we introduce the notion of protocol to describe the agents' illocutory moves, and the notion of scene to link these moves to the evolution of beliefs and actions.

In developing this formalism we have tried to keep it as standard as possible, therefore we have attempted to use a model-theoretic type of framework. Nevertheless, this formalism has been developed with actual implementation in mind, thus, given a decidable contextual framework, agent illocutions are to be interpreted by actual messages among agents --messages that are exchanged through a conventional network and, in order to be admissible, need to sumbit to the stated dialogical protocol-- , effective procedures are the actual interpretation of most illocutory acts, and other actual real-world entities --such as legal proof-of-solvency-- are supposed to avail linguistic entities.

Our formalism "Towards Layered Dialogical Agents" is related to other formal treatments of speech acts and computational logic. Specifically in the idea of postulating an effective procedure standard interpretation for actions we are close to [Lukaszewicz & Madalinska-Bugaj, Guinchinglia et al.]. It is different from most classical AI action-situation approaches [e.g. McCarthy & Hayes, Lifschitz, Hoare, Van Bentham] in two distinctive features: first, the emphasis we put on the intensional content and entailments of agents' illocutory exchanges --where we are closer to the formal semantics of [Singh]-- and, second, the structured protocols these exchanges are subject to --and in this respect we are close to some of the conversational issues raised by Winograd and Flores [Winograd & Flores] and their computational implementations [Medina Mora et al.]. By so doing, we have come to address many of the issues that have been considered by the more philosophical treatment of dialogues, such as [Hamblin, Rescher, Hintikka], but ours has a distinctively computational flavor.


This work has been partially supported by the European HCM project VIM, number PL93-0186, contract CHRX-CT93-0401; the Spanish CICYT project SMASH, TIC96-1038-C04001; the Italian-Spanish collaboration project CNR-CSIC, 132.06.1; the Mexican CONACYT grant [69068-7245], and the British EPSRC grant GR/K27957.

We are gratefully indebted to Xavier Márquez, manager of the Blanes fishermen's cooperative market (Blanes, Girona, Catalunya) for his illuminating explanations. We also want to acknowledge Andreas Kind and Julio García contributions.

The Fishmarket Project logo and the original webpage layout were designed by Gemma Sales.
Updated: January 8th, 1998

[ IIIA | IIIA Agents | Fishmarket Project ]