applications of intelligent agent-based systems

APPLICATIONS OF INTELLIGENT AGENT-BASED SYSTEMS Eugénio Oliveira Faculdade de Engenharia, Universidade do Porto, NIAD&R-LIACC Rua dos Bragas 4099 Porto Codex, Portugal [email protected]

1

INTRODUCTION

Since the early eighties, Distributed Artificial Intelligence emerged as a promising area for modelling and implementing distributed computational processes and entities displaying some kind of intelligence. This paper presents our perspective of the potential application of Multi-Agent Systems together with some ideas on the relevant research that makes those applications possible. A more extensive reflection over the same subject can also be found in [Oli, Fis, Ste, 99]. The structure of the article is as follows: We first give, on chapter 1, the basic definitions both for agent and multi-agent systems. Chapter 2 introduces the main methodologies that are agent related (agents’ interaction, distribution and learning). Chapter 3 puts forward several application domains and possible solutions where agent-based systems have proved to be useful. Chapter 4 gives a short conclusion. Agent The concept of agency is being now broadly used not only as a model for computer programming units displaying a certain kind of characteristics but also in a more abstract and general way, as a new metaphor for the analysis, specification and implementation of complex software systems. Many authors have given definitions of an agent [Jen, Syc, Woo 98]however a main distinction is that an agent, unlike other programs, must simultaneously have at least the following main features: - It perceives the world where it is situated. - It has the capability of interacting with other agents. - It is pro-active in the sense that it may take the initiative and persistently pursues its own goals. Different kinds of agents can be characterised, and other sophisticated features are usually needed to make them act intelligently in their respective environment. Nevertheless, it is usually assumed that the core of an agent includes the three characteristics mentioned above. Multi-Agent Systems Although, in many cases, agents can act separately to solve a particular problem, it often happens that a complete system

made of several different agents has to be designed to cope with a complex problem involving either distributed data, knowledge, or control. A multi-agent systems can therefore be defined as a collection of, possibly heterogeneous, computational entities, having their own problem solving capabilities and which are able to interact among them in order to reach an overall goal. Multi-Agent Systems’ approach implies the use of a methodology enabling the successful resolution of several problems such as: the specification formalism, the communication protocols, the agents’ coordination, the computation efficiency, the implementation tools and the verification methods. We are not going to analyse all these problems in this paper. Instead, we shall pick out some important points, which are critical in order to make MAS useful and applicable in the present. There are several possible architectures for designing and implementing agents, ranging from one extreme of simple reactive agents [Bro 86] to the other extreme of heavily cognitive and deliberative agents [Wit 92], [Jen, Syc, Wool 98] ]. Each one of these pure paradigms can be useful into very distinct application domains. Robots’ teams for accomplishing very simple missions is seen as a good example for the application of reactive agents while co-operative expert systems for management and control of complex scenarios either in distributed networks or in manufacturing are seen as good applications for cognitive agents. There are also a number of hybrid architectures rising from the explicit need of particular applications, to which both former approaches are not adequate, as it is the case with more elaborate robotic tasks. [Fir 92] [Oli 97] [Nev, Oli 97]. An important open issue in hybrid multi-agent systems (as well as for the agent itself) is what we can call the problem related with the “schizoid” syndrome [Oli 97]. The source of this syndrome is the fact that several different agents with different capabilities, architectures and response times, may be competing for controlling a system (as it is the case in the control architecture of a mobile robot). Since we do not always want to establish an a priori hierarchy to resolve conflicts, reactive agents may be proposing all the time actions that

contradict or supersede more elaborate deliberations coming from cognitive agents working in parallel A flexible, hybrid architecture which is able to decide when to pay attention to one or another kind of the proposed actions, or even if they should or should not be combined, leading to a kind of fuzzy control, is still a matter of investigation [Nev, Oli 97] . This problem of conflict resolution is somehow avoided, in most hybrid architectures, by using layering as an agent organisation principle. This principle has some advantages with respect to conceptualisation (modularity), robustness (fault tolerance, easy debugging), efficiency (agents on different levels run in parallel), clearness (clear separation between reactive and deliberative agents). This is the case with the RAP system [Fir 92] for robot control and with InteRRaP [Mul 96]. InteRRaP can be seen as a general agent-based layered system where control is shifted bottom up from one layer to the other (from reactive to more cognitive agents that propose plans and co-operation) whenever one layer is not competent enough to deal with the current situation. However, this solution is not always acceptable due to its inflexibility coming from the fact that a kind of serialisation of competencies has to be provided since the beginning by the designer. Besides control, also knowledge is seen as a key issue to characterise an agent’s architecture. The question of what kind of knowledge has to be included in each agent’s mental state, as well as how it has to be represented, has lead to the, now very popular, BDI (Believes, Desires, Intentions) architecture [Woo, Jen 95]. These mental categories, which have already been enhanced by other concepts, like goals, commitments, and plans, are still far from being well defined for generic situations.

2

AGENTS’ DISTRIBUTION AND INTERACTION

Multi-agent Systems has been used as a framework for tackling complex problems where task distribution is of primary importance. According to Durfee [Dur 88], task distribution can be recommended according to the following criteria: - to avoid overloading of critical resources, - to assign tasks according to appropriate agents competencies - to enable possible further sub-decomposition by some agents - to minimise communications through appropriate clustering of agents A variety of different mechanisms for co-ordinating agents in a multi-agent system are already available: - Contract Net Protocol [Dav, Smi 83] which proposes episodic rounds of inter-communication acts (announcements, bids, award messages). It is a simple and widely used protocol that, on one hand, does not affect too much the system responsiveness but, on the other hand, neglects possible strategic reasoning capabilities of the responding agents. The Contract Net Protocol is mainly applicable to well-defined coarse-grained task decomposition. - Multi-agent planning which implies that all agents have planning capabilities and, each one of them, takes into account other agent’s actions and constraints. Local plans can be communicated to other nodes in the network in order to reach a global plan

-

accepted by all agents (that are willing to collaborate) before the actual action execution. This is the case with Partial Global Planning (PGP) first proposed by Durfee [Dur 88] and then enhanced further by Decker in TAEMS [Kei, Les 95] to ensure more general applicability and ability to deal with real-time problems. Computational market-based mechanisms which can be designed to enhance the adaptation, robustness, and flexibility of multi-agent systems. Within a market, agents representing (or providing) services and/or competencies available in the multi-agent system compete to perform tasks leading to the satisfaction of their own individual objectives as well as to a possible overall system’s goal.

Market-based co-ordination strategies usually use auctioninspired protocols not only to empower, through computers and Internet, the real market activities, but also to facilitate distribution of task and resources allocation. Wellman [Wel 96]advocates market-based mechanisms for flexible management of information systems, claiming that a market’s ability to rapidly disseminate (through price variation) changes in the scarcity value of remaining resources, minimises communication and avoids global control. The practical reasoning paradigm that can be found in BDIbased agent architectures assumes a joint intention mental state enforcing the agents to commit to specific (eventually persistent) goals. Commitments are usually related to the execution of tasks, allocation of resources or exchanges of partial or final results. Commitments may or even should be established through negotiation. Negotiation is the process for self-interested autonomous agents to solve conflicts under bounded resources Main approaches for agent decision making while negotiating can be based on game theory. Agents using payoff matrixes to represent common knowledge, exchange their offers until an acceptable deal is reached. Variations of this principle can be found, for example, in the PERSUADER system [Syc 90]. In Open computing system like most multi-agent systems, paying attention to dynamics is of great importance. New agents can all the time arrive to or leave the environment, often the internet, thus influencing the external conditions where other agents work. Both specialised agents for mediating newcomers as well as specialised languages to make them understand each other are crucial. Several attempts to impose inter-agents standard languages using speech act based theory are on the way. KQML [Fin, Fri, Mck, McE 94] and ACL recently proposed by FIPA (Foundation for Intelligent Physical Agents) are the most promising candidates. These languages besides to be based on appropriate performatives of the speechact theory, clearly distinguish between several message levels: the content, the container and the identification. Learning For most application tasks, even in environments appearing simple, it is difficult (or even impossible) to determine the behaviour of a multi-agent system a priori - that is, at the time of its design and prior to its application. This would require, for instance, that the designer knows in advance which environmental requirements will occur in the future, which agents will be available at the considered time, and plan their interaction in response to these requirements.

In dynamic scenarios (like markets) where agents can appear and disappear, where strategies can be continuously modified, and where external conditions can be unexpectedly changed, it is extremely difficult, sometimes impossible, to define a priori the best strategies to be used by the agents. Consequently, agents need to be endowed with the ability to quickly adapt their behaviour in accordance with dynamic changes. Such capabilities should also include responses to changes in the other agents’ behaviour. Whether these reactions can be dealt with in the same manner as any other environmental factor or whether they should be dealt with in a special way, is an open issue. This controversy can be seen as a debate between modelbased learning (to compute the agent’s next action by taking into account a model of its acquaintances) and direct learning (to directly learn, from past experience, the expected utility of the agent’s action in a given state) [Par, Ung 97]. Once the dynamics of the scenario (like for instance in a market-place) is considered, the difficulties associated with learning increase still further. This is because an agent also has to determine how other agents may also change their strategies through some learning capability. This possibility leads to another dichotomy between what can be called “myopic learning agents” [Vid, Dur 97] that use a simple, short-term learning model and strategic learning agents that consider a long-term model including, in their own model, the learning process of other agents. Our work [Oli, Fon, Jen 99] and [Car, Oli, Sch 99] follows the former approach and aims to endow agents in electronic markets with simple and effective learning capabilities. Such agents need to adapt to the changing conditions whether they are buyers or whether they are sellers. We believe this feature is essential if agents are to be successful. In our experiments, agents offering their own resources (time availability) in response to the announcement of future tasks to be executed, apply a reinforcement learning algorithm to determine the most appropriate price to bid under current market conditions (including the amount of possible future work). Therefore, we propose an on-line, continuous learning mechanism that is especially adapted for agents to learn how to behave when negotiating resources (goods or services). The aim of this work is then to give a contribution to enhance, through the agents’ adaptation, their performance in the electronic market environment. Agent learning can be further sub-divided into isolated learning and interactive learning. In the first case, classical inductive algorithms can be used and the results either be used as property of a single agent or to be made available to the other agents in the community. On the other hand, interactive learning in a multi-agent system, the agents communicate in order to learn how to perform joint actions or how to change their own believes according to other ones believes. Learning in a multi-agent system environment does not imply that the agents are purely co-operative. In many cases, like electronic commerce, agents may learn on how to behave in the presence of other specific agents and, still, stay competitive in the market.

3

MAS APPLICATIONS

With the progressive importance of distributed computing, agents and the agency concept is becoming inherent to the design and implementation of any complex program.

Solutions for those complex problems are now including the paradigms of distribution, decentralisation, flexible interaction, openness and adaptation. We will here just mention some examples of specific domains where the agent-based solution proved to be most appropriate. Softbots Differently to physical agents, which control a physical body in a physical environment, like robots, softbots are software agents living in virtual environments. They usually act as delegates of a human or an organisation trying to fulfil their ultimate goals (or intentions). Information agents, in general, are computational software systems that have access to multiple, heterogeneous and geographically distributed information sources as in the Internet or corporate Intranets. The main task of information agents is to perform active searches for relevant information in non-local domains on behalf of their users or other agents Applications of intelligent information agents range form relatively simple in-house information systems, through largescale multi-database systems, to the visionary Infosphere in the Internet. Commercial aspects of information gathering on the Internet are becoming more and more relevant: for example, agents may act as intermediaries supplying the relevant information according to the users needs. Information retrieval softbots are successful examples of intelligent agents “leaving” in the internet seeking for relevant web pages, containing meaningful information seen as important for the user. Amalthaea [Mou, Mae 98]. Intelligent Manufacturing Systems The design and control of intelligent manufacturing systems is an important goal for Multi-agent system that has deeply influenced holonic manufacturing solutions. [Dee 94] [Fis 98] [Bus 96]. Within a CIM system we can identify different layers of abstraction: workflow management, shop floor control, and autonomous control systems. Parunak presented YAMS [Van 87] as one of the first approaches models to design a flexible manufacturing system (FMS) with a DAI approach. The main idea of YAMS is to use the contract net protocol [Dav, Smi 83] for task allocation in the FMS. [Ow, Smi, How 88] and [But, Oht 92] as well as to use the contract net model for job shop scheduling. How task allocation can be done using a reactive scheduling approach was described in [Hah, Lev 94] [Fis 94] and proposed a completely decentralised model for job shop scheduling which is able to produce better results than the pure contract net protocol because planning is done with some lookahead. Distributed multi-agent systems paradigm as well as interaction policies are now accepted as appropriate methods for modelling and controlling the several different levels (from the control of the material flow to general task allocation) of manufacturing systems. Contract net is seen as the basic principle for interaction that has to be further developed to meet the requirements of sophisticated negotiation between agents representing self-motivated resources.

Virtual Organisations One of the areas where improvements are starting to be especially visible, is the information technology for supporting business practices. Information technology can be helpful in providing infrastructures like: - Computer and communications networks and associated intra, inter or local area networks; - Large and distributed data bases; - Agent-based programming. We may define Virtual Organisation (VO) (identified as Virtual Enterprise by many authors) as a temporary network of independent companies having several different roles (suppliers, customers, sellers and buyers) connected through a communication network in order to share skills and competencies to access new markets. Due to their greater flexibility and agility, VO is capable of responding to rapidly changing requirements. To external partners this network of independent companies act, however, as a single corporation. The corporation refuses an institutionalisation, e.g., by central offices; instead, the cooperation is managed by using information and communication technologies [Fis, Mul, Pis 96] [Arn, Fai, Har, Sie 95]. Workflow management deals with the specification and execution of business processes. General process definitions include activities to be performed, their control flow, and data exchange. They also comprise organisational roles of persons and software components that are permitted to perform activities. Policies, which describe the organisational environment, complete a process definition [Mer, Lie, Lam 97]. Workflow management tools seem to be ideal means to realise new organisational structures in enterprises of the future The VO life cycle can be decomposed in four phases [Fis 96] [Fis, Mul, Hei, Sch 96]: - Identification of Needs: conceptual design of the VO. The VO description is product driven. - Partners selection: rational selection of the individual organisations that will compose the VO, based in its knowledge, skills, resources and availability. - Operation: control and monitoring of processes between partners, including resolution of possible conflicts, and VO reconfiguration due to partial failures. - Dissolution: breaking up the VO, and distribution of the obtained profits. Virtual organisations can be characterised along several different axes: - Scope of the activities: either single or multiple line - Range of influences: local, national, regional, global - Clients: organisation working for specialised clients or for the global market If we follow M. Porter’s organisation strategic analysis we may ask how is VO placed to answer the challenges of the market: 1. What is the level of competition between organisations offering similar products? A: VO introduces at this level a new paradigm: the possibility of easily creating temporary alliances with previous competitors as long as it is profitable for all of them. 2. How can suppliers as well as customers influence and possibly impose a pressure on the Organisation?

A: VO gives the possibility for customers and suppliers to be on the same side and create alliances to keep useful common policies (concerning pricing, quality, etc). 3. How does the organisation evaluate the risk of introducing new products in the market? A: It is hopefully that VO can early detect threads and dangers coming from the bad market reactions to both the introducing of new or the substitution of old products. According to the economists’ analysis, and comparing both advantages and disadvantages, we tend to conclude that the strengths and opportunities are much stronger than the possible weaknesses and threats one and, therefore, VO gives a chance to get more competitive advantage on the market. Electronic Commerce Commerce is the way customers and suppliers meet at a certain place and a certain time in order to announce buying and selling intentions that eventually match and successfully start business transactions. Due to innovations in information and communication technologies of the past years, time and space restrictions have been weakened and, therefore, those business transactions became easier. A business transaction can be defined as a set of interaction processes between participants playing different roles customers, suppliers and eventually intermediaries. This transaction is considered as complete when a trading agreement is made between customer and supplier, and the exchange of products or services takes place. The life cycle of business transactions can be decomposed in three fundamental stages: information, negotiation and exchange [Wel, Mac, Sug]: i) Information The market environment is described, and relevant information about participants has to be made available. This information includes available products or services, their respective specifications, suppliers and delivery terms, as well as identification of potential customers. Notice that customers can be stimulated through product information. This stage result in a list of potential partners in the market and their respective offers and demands. ii) Negotiation It consists in a decision-making process where involved parties (customers and suppliers) jointly reach a space of possible solutions. Protocols and strategies must define such interaction. This stage results in a matching pair of supply and demand. In traditional retail markets, prices and other aspects of the transaction may be fixed. In other markets, however, price value as well as other factors involved on the deal are subject to agreement between participants. iii) Exchange In this stage the transaction is done, goods or services are exchanged for (secure) payments. Optionally, this stage includes after sales service and the customer satisfaction evaluation. Traditional commerce concerns the establishment of business transactions where participants meet personally in physical places.

Today, the convergence of new advanced computing and telecommunications technologies is radically changing the traditional commerce. Due to the rapid growth of network communications, commerce participants are now facing a new innovative and challenging way of doing business: the socalled Electronic Commerce. Electronic Commerce (EC) uses the information and communication technology enabling the electronic support and, optionally, the business transaction automation, at least up to the completion of the negotiation stage. Systems using agent technology can be applied to the different stages of the Consumer Buying Behaviour model, as explained in [Gut, Mou, Mae 98]. In particular, when applying to the negotiation process, agents enable new types of transactions, where prices and other transaction terms need no longer to be fixed. A typical case of negotiation process is the auction. Negotiation on the Internet often amounts to one party (typically the seller) presenting a take-it-or-leave-it proposal (e.g., a sale price). Auctions represent a more general approach to look for an appropriate price, admitting a range of negotiation protocols [Wur, Wel, Wal 98]. These auction-based negotiation protocols include the English auction, the Dutch auction, the First-price Sealed Bid and the Vickrey auction. These auction-based protocols are called single-sided mechanisms, because bidders are all buyers or all sellers, and they also include an auctioneer. Double-sided auctions admit multiple buyers and sellers at once, like the continuous double auction. The number of on-line auctions is increasing, and in many cases they consist on augmentations to retail sites with retailers playing the roles of both auctioneer and seller [Gut, Mae 98]. There are a number of examples of auction-based systems that are in use for electronic commerce. AuctionBot is an auction server where software agents can be created to participate in several types of auctions. Kasbah is a Web-based system where users create autonomous agents that buy and sell goods on their behalf [Cha, Mae 96]. It is based on continuous double auction mechanisms. In the Kasbah marketplace, buying and selling agents interact and compete simultaneously, making price bids to their counterparts. The agents have a function that changes the price they are willing to propose over time. A different negotiation approach involves negotiating over multiple terms of a transaction, unlike the typical way of auction negotiation, that only considers the price of the good. Tete-a-Tete is a system where agents co-operatively negotiate in this way. It intends to provide merchant differentiation through value-added services, like warranties, delivery times, etc. [Gut, Mou, Mae 98]. Tete-a-Tete employs a combination of MAUT and DCSP techniques [Gut, Mae 98b]. Following this more general approach, in [Far, Sie, Jen 98] negotiation is defined as a process by which a joint decision is made by two or more parties. The parties first verbalise contradictory demands and then move towards agreement by a process of concession making or search for new alternatives. In that paper, a many parties, many issues negotiation model was adopted, that is, multilateral negotiations (like a continuous double auction) about a set of issues (the transaction terms). Several negotiation tactics were tested, and a negotiation strategy model is explained.

SMACE [Car, Oli, Sch 99] aims to combine those tactics in a dynamic way, in order to endow adaptive market agents with strategies (appropriate ways of selecting combinations of tactics) and to compare their performance with agents that do not change their combination of tactics. Since the marketplace is a dynamic environment, adaptive agents are expected to be able to benefit from changing conditions, therefore taking advantage over others. A test bed has been created to provide the interaction process between market agents with different capabilities. Entertainment There are a large number of games available in which animated characters face challenges in a virtual world. Fighting and shooting games are the most prominent examples. However, there are also adventures available in which the actions of the characters are not so cruel. Grand and Cliff [Gra, Cli 98] brought together agent technologies and concepts from Biology and designed the game Creatures. Creatures provides a rich, simulated environment containing a number of synthetic agents that a user can interact with. Interactive theatre and cinema are other application examples, which are particularly demanding with respect to the abilities of the participating agents. A number of projects have been set up to investigate the development of such agents [Tak, Nis, Mor, Hat 97] [Hay 95] . Traffic Management Car Traffic management is a complex problem that can be seen as a good example of an inherent distributed problem. Traditional approaches to this problem usually fall into static and/or centralised solutions. The static solutions have the big disadvantage of being not flexible enough to appropriately respond to frequent traffic changes. On the other hand, reactive agents solutions have the advantage of being totally independent and autonomous The centralised solutions have the advantage of looking more "intelligent" when making decisions because they can use more information and more complex strategies. The big problem is that the centralised decision limits the scalability and expandability of the systems. Broadening the area under consideration may lead to bottlenecks at the decision stage. Communication between traffic lights, controllers and controlling centre makes all the system heavy and decreases its efficiency. Several different AI-based strategies have been proposed, in the past, for the traffic management problem [Bie 94]. TraMAS [Fer, Oli 99] approach to traffic management is based on a Multi-Agent System (MAS) where agents are simple and react to both the conditions of the surrounding environment (typically a crossroad) and to the messages some of the adjacent agents (also representing crossroads) are exchanging. Because we have in most traffic telematics applications a link between the physical entities and the workflow of an organisation, we have in most of these settings actually both types of agents pure software agents and agents with Gestalt. Traffic telematics is therefore a typical example of an application domain with a hybrid agent structure.

Network Management Some of the first applications of multi-agent systems were related with network management. The obvious idea was to apply to an already inherent distributed and decentralised problem an equally distributed and decentralised agents-based solution. ARCHON [Wit 92] was both a generic platform and a methodology to build up multi-agent systems integrating heterogeneous expert systems like agents. All the agents, besides to maintain their own specific goals, were able to jointly work, co-operatively, perform an overall system’s goal. ARCHON was used for managing and controlling a big electric energy distribution network. Other applications range from robotics cell control to experiments control in CERN.

4

CONCLUSIONS

Sophistication, heterogeneity, distribution and decentralisation are characteristics of many application domains that require appropriate and effective solutions. This paper aimed at presenting agents and multi-agent based systems as possible candidates for a new paradigm of designing and implementing many of old as well as new applications seen as important in the current information society. Although many of the examples here referred already are in practice, we also should stress out that several issues still deserve more attention and further investigation is needed. Virtual Organisations, Electronic Commerce, Entertainment industry, Intelligent Manufacturing Systems will benefit from future developments in the fields of automatic learning, more cognitive agents’ architectures or enforcement of social laws.

REFERENCES [Arn, Fai, Har, Sie 95] O. Arnold, W. Faisst, M. Härtling, and P. Sieber, Virtuelle Unternehmen als Unternehmenstyp der Zukunft? in: Handbuch der modernen Datenverarbeitung Theorie und Praxis der Wirtschaftsinformatik, vol. 185 (Hüthig-Verlag, Heidelberg, 1995 [Bie 94] M. Bielli, G. Ambrosino and M. Boero (Eds.) “Artificial Intelligence Application to Traffic Engineering”, VSP, Utrecht, ISBN 90-6764-171-5, 1994 [Bro 86] R. Brooks, A robust layered control system for a mobile robot, IEEE Journal of Robotics and Automation, V.RA-2 (1), 1986.

[Cha, Mae 96] A. Chavez and P. Maes. Kasbah: An Agent Marketplace for Buying and Selling Goods. Proceedings of the First International Conference on the Practical Application of Intelligent Agents and Multi-Agent Technology (PAAM'96). London, UK, April 1996. [Car, Oli, Sch 99]H.cardoso, E.Oliveira, M.Schaefer, A Multiagent System for Electronic Commerce enabling adaptive strategic behaviours, submitted to EPIA’99, Portugal. [Dav, Smi 83] R. Davis and R. G. Smith, Negotiation as a metaphor for distributed problem solving, Artificial Intelligence 20 (1983) 63-109. [Dee 94]S. M. Deen, A Cooperation Framework for Holonic Interactions in Manufacturing, in: Proc. of the 2nd int. working conference on Cooperating Knowledge Based Systems, pages 103-124, (DAKE Centre, Universtity of Keele, June 1994 [Dur 88] E. Durfee, Coordination of Distibuted Problem Solvers, Kluwer A.P., Boston, 1988. [Fai 97] Wolfgang Faisst. Information Technology as an Enabler of Virtual Enterprises: A Life-Cycle-Oriented Description. Proceedings of the European Conference on Virtual Enterprises and Networked Solutions, Paderborn, Germany, April 1997. [Far, Sie, Jen 98] P. Faratin, C. Sierra, and N. R. Jennings: Negotiation Decision Functions for Autonomous Agents. Int. Journal of Robotics and Autonomous Systems 24 (3-4) 159-182. 1998. [Fer, Oli 99] J.M.Fernandes, E.Oliveira, TraMas: Traffic Control through Emergent Behaviour. Proceedings of PAAM 99, London. [Fin, Fri, Mck, McE 94] T. Finin, R. Fritzson, D.Mckey and R.McEntire, KQML as an agent communication language, in: Proc. of the Int. Conf. on Information and Knowledge Management, ACM Press, New York, 1994. [Fir 92] R.J. Firby, Building symbolic primitives with continuous control routines, in: Proc. 1st Int. Conf. on AI and Planning Systems (Morgan Kaufmann, CA, 1992) [Fis 98] K. Fischer, An Agent-based approach to holonic manufacturing system, in Intelligent Systems fo Manufacturing-Multiagent Systems and Virtual Organisations (Kluwer Academic Publisher, 1998.

[Bus 96] S. Bussmann. A Multi-Agent Approach to Dynamic, Adaptive Scheduling of Material Flow. In J. W. Perram, J.-P. Müller (eds.): Distributed Software Agents and Applications (MAAMAW'94), LNAI 1069, (SpringerVerlag, 1996) 191-205.

[Fis, Mul, Hei, Sch 96] Klaus Fischer, Jorg P. Muller, Ingo Heimig, A.-W. Scheer. Intelligent Agents in Virtual Enterprises. Proceedings of the First International Conference on the Practical Application of Intelligent Agents and Multi Agent Technology, London, UK, April 1996

[But, Oht 92] . Butler and H. Ohtsubo, ADDYMS: Architecture for Distributed Dynamic Manufacturing Scheduling, in: A. Famili, D. S. Nau, S. H. Kim, eds., Artificial Intelligence Applications in Manufacturing (AAAI Press/MIT Press, 1992) 199-213.

[Fis, Mul, Pis 96] K. Fischer, J. P. Müller, M. Pischel, A pragmatic BDI architecture, in: M. Wooldridge, J. P. Müller and M. Tambe, eds., Intelligent Agents II: Agent Theories, Architectures and Languages, LNAI 1037, Springer-Verlag, 1996.

[Fis 94] K. Fischer, Knowledge-based Reactive Scheduling in a Flexible Manufacturing System, in: E. Szelke and R. M. Kerr, eds., Knowledge-Based Reactive Scheduling (North-Holland, 1994.

[Oli, Fis, Ste 99] Oliveira E., Fischer K., Stepankova O., Multi-Agent Systems, Which research for which applications, in Robotics and Autonomous Systems, V28, N.1-2, pgs. 1-16. Elsevier, 1999.

[Gra, Cli 98] S. Grand and D. Cliff, Creatures: Entertainment software agents with artificial life, Autonomous Agents and Multi-Agent Systems 1(2) 1998.

[Oli, Fon, Jen 99] Oliveira E., Fonseca J.M., Jennings N., Learning to be competitive in the market, in Proceedings of AAAI’99 Workshop on Negotiation, ed. S.Sen, Orlando 1999.

[Gut, Mae 98] R. Guttman and P. Maes: Agent-mediated Integrative Negotiation for Retail Electronic Commerce. Proceedings of the Workshop on Agent Mediated Electronic Trading (AMET'98). May 1998.

[Oli, Fon, Gar 1997] Oliveira E., Fonseca J.M., Garção A., Maciv: a DAI based resource management system, in International Journal on Applied Artificial Intelligence, V.11, N.6, pgs. 525-550, Taylor & Francis, 1997.

[Gut, Mae 98b] R. Guttman and P. Maes: Cooperative vs. Competitive Multi-Agent Negotiations in Retail Electronic Commerce. To appear, Proceedings of the Second International Workshop on Cooperative Information Agents (CIA'98). Paris, France, July 1998. [Gut, Mou, Mae 98] R. Guttman, A. Moukas, and P. Maes: Agent-mediated Electronic Commerce: A Survey. Knowledge Engineering Review, June 1998. [Hah, Lev 94] S. Hahndel and P. Levi, Optimizing distributed production planning. In Proceedings of the 2nd International Conference on Intelligent Systems Engineering ´94 (Hamburg, September 1994) 419-424 [Hay 95]B. Hayes-Roth, Agents on stage: Advancing the state of the art in AI, in: Proceedings of the Fourteenth International Joint Conference on Artificial Intelligence (IJCAI-95), Montréal, Québec, Canada, August 1995, 967-971 [Jen, Syc, Woo 98] N. Jennings, K. Sycara, M. Wooldridge, A road map of Agent research and development, in: Agents and Multi-Agent Systems 1(1) (1998) [Kei, Les 95] K.Decker and V.Lesser, Designing a family of coordination algorithms, Proc. of the 1st ICMAS, S. Francisco 1995. [Mer, Lie, Lam 97] M. Merz, B. Liebermann, and W. Lamersdorf, Using Mobile Agents to Support InterOrganisational Workflow Management, Applied Artificial Intelligence 11(6) 551-572, 1997. [Mou, Mae 98] Moukas, A. and Maes,P. Amalthaea: An evolving multi-agent information filtering and discovering system for the WWW, in Autonomous Agents and Multi-Agent Systems, V.1 N.1, 1998 [Mul 96]J. P. Müller, The design of Intelligent Agents: A layered approach, LNAI 1177 (Springer Verlag, 1996) [Nev, Oli 97] M. Neves and E.Oliveira, A control architecture for an autonomous mobile robot, in Proc. of 1st Int. Conf. on Autonomous Agens (Marina del Rey, CA, 1997) [Oli 97] E. Oliveira, Robots as responsible Agents, Proceedings of IEEE Systems Man and Cybernetics (Orlando, 1997)

[Ow, Smi, How 88] P. S. Ow, S.F. Smith, and R. Howie, A Cooperative Scheduling System, in: M.D. Oliff (ed.), Expert System and Intelligent Manufacturing 43-56, 1988. [Par, Ung 97] Parkes D., Ungar L., “Learning and adaptation in Multiagent Systems”, Proceedings of the AAAI’97 Workshop on Multi-agent Learning, Providence, Rhode Island, 1997. [Syc 90] K. Sycara, Negotiation planning: An AI approach, European Journal of Operational Research 46 (1990). [Tak, Nis, Mor, Hat 97] K. Takahashi, Y. Nishibe, I. Morihara, and F. Hattori, Intelligent pages: Collecting shop and service information with software agents, Applied Artificial Intelligence 11(6) (1997) 489-500. [Van 87]H. Van Dyke Parunak, Manufacturing experience with the contract net, in: M. N. Huhns, ed., Distributed Artificial Intelligence, Morgan Kaufmann Publishers, 285-310, 1987. [Vid, Dur 97] Vidal J., Durfee E.. “Agents learning about agents: a Framework and Analysis”, in Proceedings of the AAAI’97 Workshop on Multi-agent Learning, Providence, Rhode Island, 1997. [Wel 96]M. Wellman , A computational market model for distributed configuration design, AI for Engineering, Design and Manufacturing 9, 1995. [Wel, Mac, Sug] Michael P. Wellman, Jeffrey K. MacKieMason, Sugih Jamin. Market-Based Adaptive Architectures for Information Survivability. Excerpts from Proposal Submitted to DARPA/ITO [Wit 92] Archon an architecture for multi-agent systems, ed. Thies Wittig, Ellis Horwood, 1992. [Woo, Jen 95] M. Wooldridge, N. Jennings, Intelligent Agents: Theory and Practice, The Knowledge Engineering Review 10(2) (1995) 115-152. [Wur, Wel, Wal 98] PR Wurman, MP Wellman, and WE Walsh: The Michigan Internet AuctionBot: A configurable auction server for human and software agents. Second International Conference on Autonomous Agents, May 1998.