Mar 13, 2001 - many researcher propagate to plan it within design and analysis of specific ... production engineering domain seems to be promising as there is not only the ..... of the post condition desired by the hosting concept. The agent's ...
In: Butala, P.; Ueda, K. (Eds.): Proceedings of the 3rd International Workshop on Emergent Synthesis (IWES ´01), March 12th-13th, 2001, Bled, Slovenia, pp.73-82.
SYNTHESIS AND ADAPTATION OF MULTIAGENT COMMUNICATION PROTOCOLS IN THE PRODUCTION ENGINEERING DOMAIN Ingo J. Timm*, Hans Kurt Toenshoff°, Otthein Herzog*, Peer-Oliver Woelk° * University of Bremen Center for Computing Technologies (TZI) Universitaetsallee 21-23 D-28359 Bremen {i.timm|herzog}@tzi.uni-bremen.de
° University of Hannover Institute of Production Engineering and Machine Tools (IFW) Schlosswender Str. 5 D-30159 Hannover {toenshoff|woelk}@ifw.uni-hannover.de
ABSTRACT The application of multiagent systems is often based on the claim that there will be an emergent behavior within these systems. To reach the emergent behavior many researcher propagate to plan it within design and analysis of specific systems, as it will not occur by chance. We are presenting a new way to adaptive agent communication protocols with respect to a possible gain of emergent behavior. Communication protocols can be generated, refined or adapted by the agents autonomously. The approach uses basic concepts of machine learning. Keywords: Emergent behavior, multiagent systems, adaptive communication protocols 1. INTRODUCTION The application of multiagent systems is often based on the claim that there will be an emergent behavior within these systems. Nowadays, many researchers are talking about the emergent effect as something that will occur if it was planned carefully within analysis and design of the system. On the other hand, however, the emergent behavior will not be achieved by pure chance. The implementation of multiagent systems in the production engineering domain seems to be promising as there is not only the hope for an emergent behavior but also the need for an adequate structural and organizational model of the real-world system. In this paper we are proposing an approach based on adaptive communication protocols to ensure the emergent behavior within a multiagent system without extensive analyzing and modeling of the system. It is embedded within the framework of enterprise agents, which are also presented in this paper.
1.1 Emergent behavior in multiagent systems Emergent behavior became one of the most important criteria for multiagent architecture evaluation from the very beginning. Most authors use this term in a metaphorical and undefined manner. In this paper a definition based on three major aspects will be used. The first aspect is focused on emergent properties as a largescale effect of locally interacting agents: “Emergent properties are often surprising because it can be hard to anticipate the full consequences of even simple forms of interaction” (Axelrod, 1997). Jacques Ferber’s view is more centralized on emergent organization: “Even societies considered as being complex such as colonies of bees or ants, should not necessarily be considered as individuals in their own right if we wish to understand their organization and the regulation and evolution phenomena prevailing there. In terms of multi-agent systems, this means that an organization can emerge from the juxtaposition of individual actions, without its being necessary to define a specific objective (an element from the assembly O) which represent such an outcome” (Ferber, 1999). The third aspect links emergence with the transition from reactive agents to deliberative ones. Doing so “the idea that intelligent behavior emerges from the interaction of various simpler behaviors” (Wooldridge, 1999) arises within this theoretical basis. A closer look on all three aspects proves that communication is the main reason for achieving global structures by local interaction, dynamic organization by simple (communication) rules, and intelligent behavior of the agent system as a whole. Such that, all three aspects are unified by dealing with communication protocols.
In: Butala, P.; Ueda, K. (Eds.): Proceedings of the 3rd International Workshop on Emergent Synthesis (IWES ´01), March 12th-13th, 2001, Bled, Slovenia, pp.73-82. Considering the emergent behavior of the multiagent which cannot be directly adapted by the agents during system, it should result from the agent’s interaction. On runtime. the one hand the cooperation and coordination is following local optimization criteria (goals) and on the other hand it has to take a joint optimization criteria 2. ENTERPRISE AGENTS (goals of the multiagent system) into account. This collaboration should lead to a global optimization of the In the framework of computer-mediated supply webs system (emergent effect). Thus, as communication is one the concept of enterprise agents has been developed at the of the main methodologies for achieving emergent TZI. Enterprise agents are integrating aspects of highly behavior, an adequate selection and configuration of complex organizations like enterprises. They are used to communication protocols is required. overcome problems in the information logistics arising with critical paths, bottlenecks and risk of failures within real-time. This approach is based on an architecture introduced in Timm (2000). 1.2 Agent communication protocols Intelligent agents represent a modern approach in Artificial Intelligence and are dealing with software entities, which can act autonomously, communicate with other agents, are goal-oriented (pro-active), and are using explicit knowledge (Weiss, 1999). They are often used for tasks, which can be hardly solved monolithically and are showing a natural distribution (Knirsch and Timm, 1999). In these domains agents gain great benefit as they are able to collaborate and solve problems in a distributed manner. This is reached by individual goal-oriented behavior of each agent and collaboration using some kind of agent communication language (Wooldridge and Jennings, 1995). Thus, the main focus lays on the communicative behavior of agents and their interaction (Conen and Neumann, 1998), (Cohen, 1997). A general and conceptual description of communication (agent communication protocol) has to integrate mutual knowledge about its topics (domain), the general process of the dialog, and its current state (Carron et al., 1999). Thus, the structure of communication protocols may be divided into a domain dependent (content, problem, topic) and a domain independent part consisting of process knowledge (methods of communication, address of partners, dialog structure). Common protocols are based on the speech-act theory for the design of the process (dialog). Knowledge sharing efforts in the framework of DARPA standardization created context description protocols for knowledge interchange in expert systems. One of the main efforts is KQML (Finin et al., 1997), (Finin et al., 1992) as a knowledge exchange protocol. This development has been widely used for multiagent systems and had great impact on the evolution of a standard agent communication language. In consequence, the agent communication language FIPA/ACL seems to be a promising base for further standardization (FIPA, 2000a). Common approaches to agent-oriented analysis and design are missing an intuitive methodology to generate and customize agent communication protocols. Furthermore, communication protocols are often defined within a static structure (e.g. based on finite automata),
Fig. 1. Architecture of enterprise agent The architecture of enterprise agents is extending the BDI approach integrating deliberative as well as reactive components. A three layer approach is used consisting of the communicator, working on a low-level realization of speech acts, the controller, determining the general agent behavior, and the executer, integrating business information systems and further information sources (cf. Figure 1). The central layer, the controller, is autonomously following preferred intentions. It is based on the BDI approach and is interacting with the communication layer performing standardized sequences of speech acts (dialogues). On the other hand it activates partial plans controlling the action sequences of the executer. For the internal selection of preferred intentions an explicit conflict management system is integrated with the conflict-based agent control (cobac) algorithm.
2.1 Communicator
In: Butala, P.; Ueda, K. (Eds.): Proceedings of the 3rd International Workshop on Emergent Synthesis (IWES ´01), March 12th-13th, 2001, Bled, Slovenia, pp.73-82. The communicator provides the only interface to the the ‘translator’, which is decrypting the inter-agent external environment, e.g. other agents. The agent message into the internal unified message format. communicator should be able to establish flexible communication channels to other agents autonomously. 2.2 Executer Being aware of possibly heterogeneous agent communication languages and representations we In contrast to the communicator layer this layer propagate a very flexible approach to agent ensures the internal communication with accessible units communication. In this framework, we are using different of its enterprise, e.g. resources. By that it is the interface agent communication languages explicitly. Therefore, we for the respective business information system or other implement generic types of communication acts within the information resources. communicator, which can be translated into different Scaling up for large applications the internal agent communication languages by partitioning and representation of an enterprise becomes itself an agent transforming the messages. system. This leads to hierarchical agent system The internal structure of the communicator can be architectures reflecting the enterprise structure. Assuming found in Figure 2. It consists of three main components: that level of heterogeneity between external agents connection, envelope and translator. exceeds the internal level significantly, the executer may be realized as a simplified communicator. 2.3 Controller
Fig. 2. Internal structure of the communicator The strict separation in three different modules should help to support the flexibility of this layer by possible substitution of each module. Therefore, interfaces between them are clearly defined. The first component ‘connection’ is responsible for the ‘lowest-level’ part of the communication, which is concerned in this approach, i.e. it uses interconnection protocols, as JavaRMI, CORBA, DCOM etc. rather than network protocols, like TCP/IP. The ‘envelope’ ensures the universal application of an agent even in heterogeneous environments. It is transforming incoming and outgoing messages into XMLstatements. Thus, they can be delivered by any service, which can transmit text strings. It also parses incoming XML statements and handles them to the 'translator' module. The ‘envelope’ contains information like used agent communication language, sender, etc. The controller layer initiates a communicative act by sending a unified message to the ‘translator’ module of the communicator layer. This module translates the unified message into the designated agent communication language (e.g. KQML). The translated message is transferred to the ‘envelope’ module. On the other hand incoming messages are transferred from the ‘envelope’ to
The controller layer determines the behavior, strategy, and state of the agent. That means an agent behaves in the way the functions and procedures in the controller decide. It can only learn from experience acknowledged in the controller. The architecture presented here is based on the deliberative agent architecture BDI (Bratman, 1987). The BDI architecture mainly consists of three mental categories: believes, desires, and intentions. The cobac approach extends the algorithm allowing for conflicts and inconsistent sets of goals within agents. 3. ADAPTIVE COMMUNICATION PROTOCOL In the first subsection we are presenting a formal model to BDI, which meets our requirements for conflictbased agent control (cobac) and open, adaptive communication protocols (oac), followed by an introduction of the cobac approach. Learning and adaptation of agents is the focus of the last subsection, which introduces open and adaptive communication. 3.1 Formalizing BDI The general shape of formalization may be explained by the following basic definitions. DEFINITION 1: A single belief is symbolized as a proposition; the set of any possible beliefs is defined as B, and the set of current beliefs as B. The agent state at the time t (St) is determined by the 3-tuple (Bt, Dt, It) of current beliefs, desires, and intentions, which are defined in the definitions 1, 5 and 6. For a better understanding we start introducing the concepts bottom-up. On the action level of the system we
In: Butala, P.; Ueda, K. (Eds.): Proceedings of the 3rd International Workshop on Emergent Synthesis (IWES ´01), March 12th-13th, 2001, Bled, Slovenia, pp.73-82. differentiate two general concepts, communicative actions tion: select: C × B → DI; select(Con, Bcon) = d* (communicator) and executive actions (executer). with d* ∈ Con. DEFINITION 2: A is the set of all generic actions. The sets of generic communicative and generic executive actions are defined as Ac ⊆ A and Ae ⊆ A, so that (Ac ∩ Ae = ∅) ⇔ (Ac ∪ Ae = A) is true. The agent has a set of predefined generic communicative action types (Ac) as well as a set of generic executive action types (Ae). The instances of these generic action types are activated either by the communicator or the executer (see Definition 2). For the next level of abstraction we unify communicative and executive behavior. Executor action sequences and communicator action sequences are integrated in the dialogue 3-tuple (Seq, Bdiag, select) (Definition 3): The dialogue concept consists of a set of available generic actions, a set of current beliefs in the framework of this dialogue, which is a subset of the current beliefs of an agent, and a mapping from both sets to one action, which e.g. implements a dynamic belief network for action selection.
The above introduced concepts, action, dialogue, and conversation, are represented by classes, which are activated by instantiation. In the case of a conversation a subset of current beliefs of the agent and in the case of a dialogue a subset of the beliefs of the activating conversation is created for them. On the basis of these concepts we define an auxiliary concept Aim, which is used by the following definitions. Aim is a subset of the set of beliefs B. It is used as a subset of the post condition desired by the hosting concept. The agent’s goals are specified as desires (see Definition 5). They consist of the set Aim and a set of alternative conversation for pursuing, i.e. achieving, Aim. Furthermore, it contains a mapping for the determination of the preferred conversation as well as the currently chosen preferred conversation. DEFINITION 5: Let D be the set of any possible desire an agent can have and D ⊆ D the current set of desires the agent has. A desire is defined as a 4-tuple (Aim, Alt, c, preferred) with Aim ⊆ B, Alt ⊆ C, c ∈ C and the mapping preferred: B × C → C; preferred(Aim, Alt) = g* with g* ∈ Alt.
DEFINITION 3: The 3-tuple (Seq, Bdiag, select) is a dialogue with Seq ⊆ A, Bdiag ⊆ B, and the mapping select: A × B → A, select( Seq, Bdiag) = a* with a*∈ Seq. The current available dialogues of an agent are symbolized with Di as a subset of all dialogues Di ⊆ DI. A dialogue has a subset of the current beliefs of the agent. They are containing necessary information about the opponent, the state of the dialogue, and support for the selection of the agent’s next action. If there is content knowledge necessary, e.g. content of a speech-act or parameters for method invocation, it is supported by the next level of abstraction, the conversations (see Definition 4). The conversation concept is slightly comparable to the concept of partial plans. It consists of pre- and postconditions, a subset of the current beliefs of an agent, a set of dialogues, and a select mapping for choosing the next dialogue to activate. The subset of beliefs can be rather complex and is representing the knowledge, which is necessary within and between the dialogues. DEFINITION 4: C is the set of all possible conversations. C ⊆ C symbolizes the set of conversations, which are implemented in an agent. A conversation c ∈ C is defined as the 5-tuple (Con, Pre, Post, Bcon, select) with Con ⊆ DI, Pre ⊂ B, Post ⊆ B, Bcon ⊆ B and the mapping select. The mapping select matches the possible dialogues Con and the believes of the conversation Bcon to a dialogue, which is designated for activa-
In the BDI as well as our approach intentions are defined as committed desires. For the activation of a desire it is necessary to choose one conversation from the set of alternatives as a basis for achievement of the goal. After this selection it is not possible to change the conversation unless by in-activation and later activation of this desire. So the definition of intention is reduced to a set of Aim, the selected conversation, and extended by a status mapping, determining the current process of Aim satisfaction (see Definition 6). DEFINITION 6: The 3-tupel (Aim, con, status) is defined as an intention with Aim ⊆ B as a set of beliefs, con ∈ C as the activated conversation and a mapping determining the completion of the Aim: status: B × C → [0..1]; status(Aim, con) = x* with x* ∈ [0..1] . The BDI algorithm consists of four steps, which are arranged in a loop and executed continuously. P symbolizes the set of perception for the agent, p ∈ P is defined as a single perception; the actions are presented in a semi-formal manner and explained in natural language. 1.
belief-revision: B × P → B
In: Butala, P.; Ueda, K. (Eds.): Proceedings of the 3rd International Workshop on Emergent Synthesis (IWES ´01), March 12th-13th, 2001, Bled, Slovenia, pp.73-82. belief-revision(Bt, p) = Bt+1 The belief revision is responsible for updating the While the main purpose of this step is to generate a set knowledge base, i.e. the beliefs, of an agent due to the of desires and committed intentions, it is also used to current perception. It initiates a new state of the transform the desires and committed intentions into a system by the generation of a new belief set (Bt+1). unique format. The elements of the resulting set are called objectives (see Definition 7). 2. options: B × I → D options(Bt+1, It) = D’t+1 DEFINITION 7: The set of objects is given by O. An The second step of the BDI algorithm creates the objective o is defined as a 7-tupel (con, Alt, Aim, options for the next action. Some approaches to BDI Pre, Post, status, priority), with con ∈ C, Alt ⊆ C, explicitly define a new set at this stage, the set of goals. This is necessary if the set of desires is allowed Aim ⊆ Post, Pre ⊂ B, Post ⊆ B and to be inconsistent. In this case the set of goals is a with the mappings status and priority as follows: consistent subset of the set of desires. We define the status: C × B → [0..1]; status(con, Aim) = x*, set D’ as a consistent subset of D. The contents of the with x* ∈ [0..1] and priority: B → ℜ+; set of goals are depending on the current state (Bt+1) priority(B) = y*, with y* ∈ ℜ+. and the set of intentions from the last step (It). 3. filter: B × D × I → I The transformation of the desires and intentions into filter(Bt+1, D’t+1, It) = It+1 objectives is based on the trans mapping (see Definition The filter mapping creates a new set of intentions for 8). this state (It+1). It is the most important step within the BDI algorithm as it determines the next course of DEFINITION 8: The trans mappings are defined for action. The selection of a new intention is called intentions and desires separately as follows: “commitment to the intention”. But the filter step is transD: D → O; transD((Aim, Alt, c, preferred)) = not only activating intentions but also retracting (d, Alt, Aim, Pred , Postd, status0, priority), with intentions, which are not appropriate any more or preferred(Aim, Alt) = d, d ∈ Alt, and cannot be achieved under the current conditions. d = (Cond, Pred, Postd, Bdcon, selectd). 4. execute-action: I → A transI: I → O: transI((Aim, d, status)) = (d, { d }, execute-action(It+1) = a with a ∈ A Pre, Post, status, priority), with d ∈ C and d = The last step in the BDI algorithm is initiating the (Con, Pre, Post, Bcon, select) previously selected action according to the current set of intentions (It+1). The application of these two mappings leads to a new set of goal, which does not have to be consistent at this The major drawback of this algorithm in the field of point of time. enterprise agents can be found in the assumption of a consistent set of desires. Modeling complex organizations 3.2.2 Filtering Using Conflict Management like enterprises often leads to definition of goals, which are inconsistent. The creation of consistent subsets (goal) The filtering process is using the conflict management for action selection is not satisfying for such propagated in this paper. The algorithm consists of three environments, as they have to realize multi-criteria main steps: optimization, which are based on competing goals. 1. Assessment of objectives 3.2 Conflict-based agent control 2. Assessment of conflicts and synergy 3. Objective selection The cobac algorithm is based on the BDI algorithm and substitutes the filtering step mainly. The structure of The assessment scheme for assessment of objectives is the cobac algorithm is presented in the following: based on the integration of a deliberative, a reactive and a relevance factor. Each of these factors is defined by a 1. Belief revision mapping from current believes of an agent and an 2. Generation of options objective to a real number. The definition of the concrete 3. Filtering using conflict management mappings is part of the customization of the agent system 4. Action execution as it determines the general weight for a designated goal and is considered within the conflict management step. As Only the steps two and three are altered and will be an intuitive approach we introduce the mappings as explained in this section. presented in Definition 9. The reactive factor is responsible for the evaluation of the objective with regard 3.2.1 Generation of Options
In: Butala, P.; Ueda, K. (Eds.): Proceedings of the 3rd International Workshop on Emergent Synthesis (IWES ´01), March 12th-13th, 2001, Bled, Slovenia, pp.73-82. to risks and expected costs when the aim is not met. DEFINITION 10: The assessment of conflicts and Furthermore, it is a required concept for the allocation of synergy is defined by two mappings, κ (conflict) resources. and σ (synergy) from the Cartesian product O × O into ℜ+. Neutrality of a pair of objectives is defined DEFINITION 9: The mapping γ is integrating the by zero used as a stopping criterion for efficiency assessment mappings ρ, δ and ω. It is defined as purposes. follows: γ: ℜ+ × ℜ+ × ℜ+ → ℜ+; κ(oi , oj) is defined as the weighted number of γ(ρ(B, oi ), δ(B, oi ), ω(B, oi )) = x, with x ∈ ℜ+ and conflicting oi post conditions and oj aims and vice oi ∈ O. The assessment mappings are all defined versa. σ(oi, oj) is defined as one plus the weighted number of supporting post conditions of oi and pre alike: ρ, δ, ω: B × O → ℜ+; {ρ|δ|ω}(B, oi ) = x, with conditions of oj and vice versa. For evaluation x ∈ ℜ+ and oi ∈ O. purposes we introduce additionally ψ as follows: ψ(oi , oj) = κ(oi , oj) / σ(oi, oj). In contrast to the assessment of reactive aspects the The concrete algorithm of these mappings is a deliberative evaluation is aiming at the estimation of the central feature of any implementation of the cobac utility of this goal. The result should determine the architecture. maximum utility within the current circumstances of an agent. The integration of the current status of a committed The last step is the objective selection process. This is intention should lead to an adequate goal pursuing the main conflict resolution phase. It incorporates an behavior, as an almost completed intention is not simply algorithm, which follows the conflict resolution steps discarded without consideration. The last aspect of the documented in Figure 3. This scheme is motivated by objective assessment is the determination of relevance. research in social psychology (van de Vliert, 1997). The relevance expressed by a dynamic weight is Different levels of conflict and synergy potential as well as depending on the state of the agent and evaluating if the the individual assessment of the objectives are leading to commitment to this desire or intention is appropriate. specific types of conflict resolution. E.g. direct fight as a strategy for disagreeable objectives and with a strong The step of assessment of conflicts and synergy dominance of the assessment of one objective remove the proposed within this architecture is dealing with conflicts other objective without changing assessment or aims of of interest within an agent (intra-agent conflict the first objective. Compromising, for example, as a management). There are different approaches to detect fruitful approach for agreeable goals will create a new and assess conflicts of interest between each pair of objective which partially subsumes the conflicting goals. objectives. The presented algorithm uses the level of Implementation of conflict resolution types is part of the inconsistency, agreement, and neutrality between two set customization process of enterprise agents to a new of desired post conditions. This assessment is extended by domain. The result is a subset of the set of objectives with level of conflicts, agreement, and neutrality for the set of updated assessment. desired post conditions from one objective and a set of pre conditions.
Fig. 3. Conflict management and resolution scheme
In: Butala, P.; Ueda, K. (Eds.): Proceedings of the 3rd International Workshop on Emergent Synthesis (IWES ´01), March 12th-13th, 2001, Bled, Slovenia, pp.73-82. general framework. It is based on a probabilistic 3.2.3 Action execution methodology, a Markov chain in analogy to dynamic The action execution consists of the activation of the belief networks (Russel and Norvig, 1994). A Markov objective o0, which is the objective with the highest model is defined by a number of dialogue states Xi , and evaluation result. If the intention of the objective o0 has propabilities prob(Xi , Xj) for a transition from state Xi to not been active, it is activated. Then the preferred state Xj. This model is highly flexible and allows adaptation, refinement and synthesis. It is easy to convert alternative conversation due to the definition of objectives standard communication protocols into this formalization is selected and the respective conversation is initialized defining all transition probabilities as zero, if no transition and started. A new dialogue is selected due to the aim of exists in the protocol, as unity, if the Xj is definitely the conversation and the conflict and synergy potential following Xi and splitting the probabilities for all possible between involved agents (inter-agent conflict branches. management). If the intention of objective o0 has been active, the active conversation and respective dialogue are continued. The analysis and design of these protocols is done with minimum effort as only required protocol structures have to be defined and initial communication protocols be 2.3 Open, adaptive communication generated. The agents are modifying their protocols As discussed above, the philosophy of this architecture autonomously during simulation and application as is to keep all executions within the frame of dialogues. follows: Moreover as discussed in the introduction, communication Adaptation The execution of existing communication is one of the main methodologies for achieving emergent protocols leads to an adjustment of the selection behavior. Therefore, an adequate selection and probabilities (transition probabilities in the model) of the configuration of communication protocols is required. next action due to prior experience with this dialogue Common approaches to agent-oriented analysis and partner, the dialogue partner's team, or the overall design are missing an intuitive methodology to generate multiagent system. This adaptation is formalized by and customize agent communication protocols. multiplication. Note that all zero transition probabilities Furthermore, communication protocols are often defined stay as they are in consequence. within a static structure, which cannot be directly adapted Refinement Within refinement an agent is extending by the agents during runtime. The agents collaborating protocols by adding new “states” Xi i.e. speech act types have to be able to adapt their protocols according to the out of a given set of basic communicative acts, e.g. dialogue partner, their own state, the multiagent system's FIPA/ACL (FIPA, 2000b). The respective transitions are state, and experience of prior communications. initialized in a straight forward manner. Another method To address this problem we propagate an approach of of refinement is to keep the states as they are, but to open, adaptive communication protocols (oac). The basic implement “new” transitions by setting zero transition idea of the architecture presented here is the incorporation probabilities to p>0, or to extinct certain transitions of the oac algorithm as the main feature of the action annulling their probabilities. Refinement is selected if an execution step. Within the oac approach three main existing protocol tends not to lead to a satisfying result extensions of classical concepts are implemented: use of according to the agent's goals and the multiagent system's communication protocols is not restricted to a given set goals. and protocols of opponents do not have to be known Synthesis The automatic generation of protocols is the (open), dialogues do not have static structures only, but desired methodology for the creation of new are flexible and can be adapted, refined and even communication protocols. The basis for this method is a synthetisized during runtime (adaptive), and conflict predefined set of dialogue “skeletons” as they are management is determining the strategy for the agent’s occurring within common communication protocols. The behavior within a concrete dialogue. first step of the synthesis is to select one of them as a core Note that in contrast to agent systems with fixed model. The necessary extension and customization of this communication protocols this approach uses a generalized rudimentary model follows the adaptation and refinement view on dialogues. In classical approaches a dialogue is steps above. Totally new protocols may be very inefficient defining an agent communication protocol, e.g. contract with respect to the special situation and the agent’s utility net protocol, Dutch-auction protocol. The concept of if adapted and refined within real future conversation. In dialogues (Definition 3) is integrating communicative as order to prevent such problems the agent may use well as executive actions. Such that agent communication recorded information of prior communication for an protocols are defined by dialogues, which consists of internal adaptation and refinement process. However, in communicative actions only. analogy to machine learning processes this can lead to The oac approach is realizing the dialogue concept some kind of overfitting of the resulting model. (see Definition 3), i.e. the select mapping, within this
In: Butala, P.; Ueda, K. (Eds.): Proceedings of the 3rd International Workshop on Emergent Synthesis (IWES ´01), March 12th-13th, 2001, Bled, Slovenia, pp.73-82. A last aspect is of great importance for this approach: agent-based systems seems to be a promising approach to In analogy to the intra-agent conflict management and bridge the gap in information logistics. using the same formalization balancing of agent goals The “IntaPS” approach uses simplified enterprise may be incorporated (inter-agent conflict management). agents for this purpose and is based on two substantial After defining the final objective o0 within the intra-agent components which link together information systems of conflict management the agent will update its estimation earlier stages of product development and the resources on of the opponent’s goal respectively the general goals of the the shop floor (cf. Fig. 4). This link is realized by system. In the next step an inter-agent conflict decentralized planning on shop floor level and by rough management takes place following exactly the rules level process planning (Toenshoff et al., 2001a). defined for the intra agent conflict management (see above and Figure 3). By this procedure the agents are capable of adapting, refining and synthesizing protocols according to specific strategies as they are indicated by the leaves of the conflict management tree (see Figure 3). Neutrality of objectives indicates that there is no conflict potential and there is no need for a change of protocols. In consequence the oac methodology leads to an easier achievement of emergent behavior than the construction of very detailed protocols in before. The oac approach is providing a dynamic communication protocol structure such that the concrete protocol emerges from the agent’s experience, current state, and opponent model. This aspect corresponds with the concept of organizational emergence (Ferber, 1999). 4. INTAPS This approach has been partially developed and is finally applied to the domain of production engineering within the research project “IntaPS - Integrated AgentBased Process Planning and Production Control” (Toenshoff et al., 2001b). The concept of a multiagent system, which models integrated process planning and production control, seems to fit very well into the specific demands within this domain. Looking at the current situation in industrial production, strong borderlines exist between process planning, production control, and scheduling systems caused both by an extreme specialization and the independent historical paths of system evolution. The traditional approach of separating planning activities (e.g. process planning) from implementing activities (e.g. production control and scheduling) results in a gap between the involved systems. This gap implies loss of time, information and in consequence loss of quality and prolonged time-to-market. In order to bridge this gap there is a strong need for new ideas integrating both worlds. Innovative fundamental concepts and methods for management and control of integrated information logistics, production scheduling and process planning are necessary. Owing to the characteristics of this domain (e.g. the natural distribution of participating entities like machines and resources, the dynamic environment and the complex interaction of the entities) the application of
Fig. 4. “IntaPS” architecture A co-operative multiagent system is implementing decentralized planning on shop-floor level. Within this architecture three different types of agents are used, resource, order, and service agent. The realization of these agents is following the enterprise agent architecture. Each relevant resource of the production system and its environment is represented by one resource agent. Consequently, agents exist for e.g. machines, transportation devices as well as staff and virtual resources like business information systems. The entirety of all resource agents represents the shop floor model of the whole production system (Toenshoff et al., 1999). Orders are represented by order agents. They pursue the goal of market-based optimization and endeavor to optimize utility functions respective individual goals, which in many cases may be competing, e.g. “rush order” versus “stock order”. Thus, individual utility functions with respect to order priority are used to evaluate possible action alternatives. They serve as the basis for the conflict management respective evaluation mappings for objectives.
In: Butala, P.; Ueda, K. (Eds.): Proceedings of the 3rd International Workshop on Emergent Synthesis (IWES ´01), March 12th-13th, 2001, Bled, Slovenia, pp.73-82. Service agents are used for human interaction, By this a highly complex communication structure transparency and maintenance purposes. They do not take emerges during runtime. advantage of the special abilities of enterprise agents. The concept has been developed in the Artificial The detailed process planning and scheduling takes Intelligence group of the University of Bremen and has place cooperatively within a marketplace. The application been refined in the framework of the “IntaPS”-project in of this approach should lead to an emergent behavior cooperation with the Institute of Production Engineering within this marketplace induced by the oac concept. Order and Machine Tools of the University of Hannover. It will agents and resource agents are interacting according to a be applied and evaluated within this project as well as a “three-phase-model”. Starting with a “negotiation phase” scenario in the transportation logistics domain. required manufacturing skills and due dates as well as capabilities and capacities are communicated. The protocols used here are already synthesized, adapted, and 7. ACKNOWLEDGEMENT refined by the communication history of the involved agents. Sequences of manufacturing operations are The presented work is part of a PhD thesis to be resulting from concrete communication dialogues completed at the TZI and the “IntaPS” project, which is following these protocols. Finally, the optimal sequence of being funded by the Deutsche Forschungsgemeinschaft manufacturing operations is accepted as detailed plan. The (DFG) within the projects He 989/5-1 and To 56/149-1 as second phase is called “verification phase” and ensures part of the Priority Research Program 1083 “Intelligent the feasibility of the detailed plan. The order agent Agents and Realistic Commercial Application Scenarios”. examines continuously within its communication protocol, Further information about the “IntaPS” project is available whether its detailed plan is executable under the current in the internet (http://www.intaps.org). conditions. Simultaneously it has to check for the estimated objectives of the resource agents involved and initiates the respective inter-agent conflict management. REFERENCES This causes order agents to analyze the consequences and to identify those parts of their detailed plans, which are Axelrod, R. (1997). The Complexity of Cooperation affected. If necessary, the order agent enters a “reAgent-Based Models of Competition and negotiation phase” and tenders parts of the detailed plan Collaboration. Princeton University Press. Princeton, for a new negotiation. Doing so, synthesis of new New Jersey. protocols can be helpful in order to establish a broad basis Bradshaw, J. M. (Ed.) (1997). Software Agents. AAAI for explicit and distributed problem solving by Press / The MIT Press. Menlo Park, California. communication. This leads to an improved alternative Bratman, M. E. (1987). Intentions, Plans and Practical plan substituting the previous one. Reason. Harvard University Press, Cambridge, Centralized rough level planning is responsible for an Massachusetts. appropriate and efficient initialization of order agents. Carron, T., Proton, H., Boissier, O. (1999). A Temporal They are created on the basis of prior protocols Agent Communication Language for Dynamic Multisynthesized, refined, and adapted by other order agents. Agent Systems. In: Multi-Agent Systems Engineering Additional geometrical or technological information about – 9th European Workshop in a Multi-Agent World, the work piece and about further organizational MAAMAW’99 in Valencia, Spain. Garijo, F. J., information related to products and orders are integrated Boman, M. (Ed.), pp. 115-127, Springer, Berlin. in the set of believes. Cohen, P. R. (1997). Communicative Actions for 5. CONCLUSION Artificial Agents. In: Bradshaw (1997), pp. 419-436. Conen, W., Neumann, G. (1998). Coordination This paper introduces enterprise agents with special Technology for Collaborative Applications. aspects as integrated conflict management and open Springer, Berlin. adaptive communication. A formal framework enhancing Ferber, J. (1999). Multi-agent systems - an introduction to the BDI architecture is provided for the internal control distributed artificial intelligence. Addison-Wesley. structure of an agent and discussed for the open, adaptive Harlow, UK. communication. In concordance with widely used Finin, T., Weber, J., Wiederhold, G., Genesereth, M., concepts of emergence for intelligent agents and Fritzson, R., McGuire, J., MacKay, D., McGuire, J., multiagent systems, the paper emphasis that emergent Shapiro, C., Pelavin, R., Beck, S. (1992). organization can be realized by flexible communication Specification of the KQML Agent Communication protocols. This approach is dominated by simple rules Language. Technical Report 92-04, Enterprise especially formalized in Markov models, but produces Integration Technologies, Inc. complex conflict dependent behavior with self-adaptation.
In: Butala, P.; Ueda, K. (Eds.): Proceedings of the 3rd International Workshop on Emergent Synthesis (IWES ´01), March 12th-13th, 2001, Bled, Slovenia, pp.73-82. Finin, T., Labrou, Y., Mayfield, J. (1997). KQML as an Agent Communication Language. In: Bradshaw (1997), pp. 291-316. FIPA (2000a). FIPA Abstract Architecture Specification. Foundation for Intelligent Physical Agents. http://www.fipa.org/specs/fipa00001/ FIPA (2000b). FIPA Interaction Protocol Library Specification. Foundation for Intelligent Physical Agents. http://www.fipa.org/specs/fipa00025/ Knirsch, P., Timm, I. J. (1999). Adaptive Multiagent Systems Applied on Temporal Logsitics Networks. In: Logistics in the Information Age – ISL 99, Florence, Italy. Muffatto, M., Pawar, K. S. pp. 213218, SGE Ditoriali. Padova, Italy. Russel, S. J., Norvig, P. (1994). Artificial Intelligence: A Modern Approach. Prentice Hall. Englewood Cliffs, New Jersey. Timm, I. J. (2000). Multiagent Architecture for D-Sifter – A modern approach to flexible information filtering in dynamic environments. TZI Technical Report No. 21/2000, University of Bremen. Bremen, Germany. Toenshoff, H. K. et al. (1999). Dynamic modelling of process planning knowledge based on a shop-floor model. In: Proc. of the 32nd CIRP Intern. Seminar on Manufacturing Systems, pp. 497-504. Leuven, Belgium. Toenshoff, H. K., Woelk, P.-O., Herzog, O., Timm, I. J. (2001a) Integrated Process Planning and Production Control. In: Proc. of the 2001 International Conference on Precision Engineering (ICPE 2001). Yokohama, Japan, to appear. Toenshoff, H. K., Woelk, P.-O., Timm, I. J., Herzog, O. (2001b). Flexible Process Planning and Production Control Using Co-Operative Agent Systems. In: Proc. of the International Conference on Competitive Manufacturing (COMA ’01). Stellenbosch, South Africa. van de Vliert, E. (1997). Complex Interpersonal Conflict Behaviour – Theoretical Frontiers. Essays in Social Psychology. Psychology Press. East Sussex, UK. Weiss, G. (Ed.) (1999). Multiagent Systems – A Modern Approach to Distributed Artificial Intelligence. The MIT Press. Cambridge, Massachusetts. Wooldridge, M.; Jennings, N. R. (1995). Intelligent Agents: Theory and Practice. In: Knowledge Engineering Review. Vol. 10. No. 2. Wooldridge, M. (1999). Intelligent Agents. In: Weiss (1999), pp. 27-78.
