M. G. Panteleev, M. A. Natey-Golenko

St.-Petersburg State Electrotechnical University “LETI”

Multi-agent systems (MAS) in which multiple problem solving agents cooperate to achieve a common objective is a rapidly emerging and promising paradigm [2,9]. The set of application domains in which agent technologies are being app lied to or investigated include air-traffic control, electronic commerce, workflow management, business process reengineering, network management, data mining, information retrieval/management, education, personal digital assistants etc. [7]. While MAS g et wide use in different areas there is a growing need of teaching specialists in relative technologies.

Agent Communication Languages (ACL) are the key component of MAS technologies. ACL allow agents to effectively communicate and exchange knowledge with other agents despite differences in hardware platforms, operating systems, architect ures, programming languages and representation and reasoning systems. KQML (Knowledge Query and Manipulation Language) is a most well known language of that kind [1,3]. It consists of primitives (called “performatives”) that express attitudes regarding t he content of the exchange and allow agents to communicate such attitudes to other agents and find other agents suitable to process their requests

A KQML message is divided into three layers: the content layer, the communication layer, and the message layer. The content layer bears the actual content of the message in a language chosen by the sender. The communication layer encod es a set of features to the message to describe the lower level communication parameters such as the identity of the sender and recipient and a unique identifier associated with the communication. The message layer encodes the message including its inten tion (by a chosen performative), the content language used, and the ontology.

Another Agent Communication Language has recently been specified by the Foundation for Intelligent Physical Agents (FIPA).

Knowledge Interchange Format (KIF), due to its rich expressive power and simple syntax, is probably the most widely used neutral message format for agent communication. KIF is a prefix version of first order predicates calculus wit h extensions to support non-monotonic reasoning and definitions [4]. The language description includes both specifications for its syntax and for its semantics. Besides first order logic expressions of facts and knowledge, KIF also supports extra-logical expressions such as those for the encoding of knowledge about knowledge and of procedures.

Every application incorporates some view of the domain it applies to. To ensure correct mutual understanding of the exchanged messages, agents must agree on the model of that part of the world about which they are exchanging inform ation with each other. In the terminology of the agent community agents must share a common ontology [5]. More formally, an ontology is a domain theory that specifies a domain-specific vocabulary of classes, properties, predicates functions, and entities , and a set of relationships that necessarily hold among those vocabulary items. A shared or common ontology refers to an explicit specification of the ontological commitments of a group of agents.

Our ultimate goal is to develop software that will provide tutoring in agent communication languages and investigation of different kinds of communication acts between deliberative intelligent agents. This software should be used for s tudying coordination and negotiation issues and different negotiation models in MAS.

In order to maximize future compatibility of our system with other ones we utilize standard technology ActiveX by Microsoft. The majority of the agent software now are implemented on JAVA: AgentBuilder® (Reticular Systems Inc.), Aglets (IBM), Intelligent Agent Factory (Bits & Pixels), LiveAgent(AgentSoftLtd.). However we believe that ActiveX technology is more preferable for future real-time agent applications.

ActiveX Agent Testbed, AxA is being developed agent building paltform based on Microsoft ActiveX technology. AxA allows to construct multiagent systems, using various development kits such as Microsoft Visual Basic, Microsoft Visual C+ +, Microsoft Visual J++, Inprise Delphi, Inprise C++ Builder etc. AxA intended primarily for research and educational purposes.

Further, ActiveX Agent Testbed is a distributed agent society also. While agents range from simple personal assistants to complex real-time research, industrial and economical agents, all have common commucation language – KQML/ KIF [4,8]. Distributed COM/ActiveX industrial standard was used as low-level transport facility. AxA provide common agent architecture, communication and interaction interfaces. AxA agents divided into two categories: application agents and base system facilitators such as agent category facilitator (described below), ontology server, resource manager. We define three layers of agent interaction, mapped to OSI layers:

• application layer — high level knowledge interchange in some language (KIF in current implementation);
• presentation & session layers — KQML;
• transport layers — DCOM/ActiveX.

As said above, AxA is based on Microsoft Distributed Component Object Model (DCOM). Microsoft defines the DCOM as a platform-independent, distributed, object-oriented, system for creating binary software components that can interac t. AxA agents are essentially ActiveX objects and communicate via interfaces. Two types of entities involved in dialog: agents and KQML messages, both ActiveX objects. AxA defines IKqmlChannel and IKqmlMessage interfaces, implemented by agents and KQML M essage objects respectively.

The sample dialog of two agents, depicted on Fig 2. Precondition: agent A must have pointer to IKqmlChannel interface of agent B before interaction. There are many ways to get it, including some name server. AxA provides proprietary Ax A Agent Category Facilitator.

• Agent A creates instance of KQML message, fills it with ontology and language names, content (message body), represented in specified language and supplies object set, referenced from message content as world objects.
• Agent A calls Process method on interface pointer it has.
• Agent B receives message and think.
• Agent B creates another instance of KQML message, fills it with answer and returns as method return value.
• Agent A reads answer calling methods on returned interface pointer.

AxA has Agent Category Facilitator (ACF) as system component and will be enriched by ontology server and resource manager.

ACF is a facilitator that provides agent naming and navigating services and works as bulletin board. ACF stores list of generalized agent categories and any agent can register itself under some categories and advertise its capabilities to society. ACF receives messages in KIF language as knowledge representation language. AxA defines two category: ontology server and resource manager, this list will extended in future.

In AxA, any agent after birth knows about one external object in world only - ACF. ACF helps agent to explore society. ACF is ActiveX object as other AxA agents and implements IKqmlChannel interface.

Exciting feature of ACF is special knowledge base. ACF operates with symbolic category names and stores interface pointers additionly. That is application agent should not to resolve given agent names to physical addresses or use some high-level ways to communicate with string names. It can invoke methods on interface pointers directly without great overhead. We extend standard KQML to support this feature.

ACF processes subset of KQML performatives include:

• Register ? register agent under specified category
• Unregister ? delete association between agent and category name.
• Ask-All ? retrieve information from knowledge base, ACF will return a list of interfaces pointers matched for requested category.

Example of possible dialogs presented below:

Agent registers itself under category “ontology-server”:

To simplify building of agent application, we include into AxA testbed additional component: Agent Communication Library (ACLB). ACLB is a library of supporting ActiveX objects and contains task-oriented negotiation models. They e ncapsulate knowledge about conducting parametrized agent-to-agent dialogs. We include into our testbed ACF Negotiation Model (CFN) and base implementation of State Oriented Domain Negotiation Model (SDN).

CFN communicates with ACF and solves two tasks: discovering partners and advertising own capabilities. CFN implements ICFNegotiator interface and provides properties, methods and events presented in Table 3:

A SDN models intended for negotiation about states, actions, goals in state oriented domains (SOD). We use SOD as defined in [11]. In brief, SOD defined as a tuple of: ordered list of agents, set of possible joint n-agent plans with ro les cost vector. A SOD world of objects could be described as a set of possible relations between agents and a set of possible weighted atomic operations on objects.

AxA agents can use various SDN models through ISDNegotiator interface, defined as follows:

AxA was designed for Windows NT (x86) and Windows 95 platforms. AxA implemented in Visual C++ and Visual Basic languages and consists of following modules:

• AxACL ? communication library, ActiveX DLL, provides base communication objects such as IKqmlMessage interface implementation and negotiation models.
• KifParser ? static library, parser of KIF forms.
• AxACF ? core of ACF, ActiveX DLL too, contains category knowledge base and KIF processor.
• AxACF_Service ? ACF shell, executable module, contains NSClient ActiveX object (allows agents to connect to ACF and interact with it) and execution GUI monitor (displays statistics and messaging activity).

AxA agent based on ActiveX technology. Structurally, agent consists of Engine, Interpretaiton Module (IM), Prediction Module (FM), Planning Module (PM), Knowledge Base (KB) and Executive ActiveX objects (EO):

All agent components are ActiveX objects.

• Engine ? core of agent, strives to move world from current state to destination state.
• Knowledge Base ? provides methods to modify and query its knowledge base.
• Executive Objects ? a set of ActiveX objects, they sense and perform actions.
• Interpretation Module ? interprets sensed data and build symbolic description of the world.
• Prediction module ? models developing of world state.
• Planning Module ? controls current situations and plans agent activity.

To teaching students agent technologies we propose to consider the next issues:

• The framework of ACL, KQML and KIF languages. Simple example of conversations as a sequences of causally-lated messages in exchanges between agents.
• Negotiation models. Different negotiation models capable of exhibiting a rich set of negotiation behaviors must be considered.
• Ontologies building and employment.
• Planning in multiagent systems.

In this paper we have described a testbed for tutoring and research in multiagent communications. Our testbed is based on powerful Microsoft’s Distributed COM/ActiveX technology and consists of reusable components, it guarantees ti ghtly integration and interoperability with widely distributed software. One of exciting features is modified KQML language, it allows to operate with COM interface pointers. Further, using modern development tools such as VisualBasic and Visual C++ simp lifies studying agent technologies.

Future work is concerned with developing ActiveX components for building more complicated deliberative intelligent agents that implement planning and prediction functions. Another future direction is concerned with applying these compl icated agents to model different negotiation strategies and cooperative planning.

We believe that the architecture and software presented in this paper will be useful both as tutoring and as research testbed for communication act studies in multi-agent systems.