1. Introduction

One of the features of recent network systems like the Internet and intranets is that they are open networks. In the open networks, their components are distributed in wide areas and managed independently. However, users cannot often use the networks efficiently because of various unexpected changes such as removal of required information to other sites or specification changes of required software.

Agent technologies are attracting attention as solutions to the problem of the unexpected changes in the open networks. In this proposal, an agent technique for the management of the unexpected changes in the open networks is described.

Summary of the technique is as follows.

• Multiagent functionalities are provided to make effective use of information distributed in the networks. These includes agent coordination techniques such as the contract net protocol and the blackboard architecture.

• Agents can move around the networks in order to handle network lines with low reliability such as wireless connection for mobile computers.

• Planning functionalities are supported to make flexible use of the above facilities with managing unexpected changes. By these functionalities, the agents automatically generate and execute the action sequences they should process on the basis of knowledge about conditions of network nodes and knowledge about their actions.

• As component units of knowledge needed for the planning, fields are supplied to make knowledge management easy.

• Users' requirements and knowledge for the planning are described in a special language for easy and efficient use of the above functionalities.

2.1 Agent Model

The following figure shows the whole architecture of the system.

The system are separated into two layers. The upper layer takes the responsibility of application level functionalities. The lower layer is called the platform layer responsible to the low level process for realization of the upper layer functionalities.

The upper layer are decomposed into components called fields which shows the extents of knowledge and actions of the agents. For each field, a knowledge manager that manages the knowledge available in the field is provided. In each field, more than one agent can exist and access the knowledge in the field through the knowledge manager. Besides the fields, blackboards are provided in this layer.

The platform layer consists of the following components.

• The field management facility manages generation, deletion and changes of the fields. Field generation proceeds as follows.

  1. The platform issues a field ID.

  2. The platform also prepares the knowledge base and the knowledge manager for the field.

• The agent management facility manages generation, deletion and changes of the agents. As for agent generation, it issues agent ID's and identifies the fields to which the agents belong. According to the agents' behaviours, it changes the agents' states such as their knowledge and conditions of communication.

• The blackboard management facility manages data of blackboards. In detail, it manages agents' access to the blackboards including read/write, deletion and changes.

  The blackboards are managed according to network nodes the blackboards exists. That is, blackboard ID's are managed by the platforms of the nodes, where the ID's can be accessed globally (for example, the blackboard whose local ID is "bb1" in the node "node1" has the global ID "bb1@node1").

• The communication management facility realizes communication necessary for inter-agent communication and agent mobility like Telescript in the networks.
  • As for inter-agent message exchange, the facility sends and receives the contents of the messages.

  • The facility provides blackboard watching daemons for inter-agent communication via the blackboards. The daemons watch events on the blackboards and notify them to remote nodes.

  • As for agent mobility, the facility can send and receive requests for remote agent generation and invocation.

• The security facility is necessary because it is problematic if the agents can communicate, migrate and act freely on the networks. This facility can limit such agents' behaviours. The facility consists of the following features.
  • Authentication limits agents' migration from other nodes and their behaviors in the nodes. The electronic signature technique can be applied.

  • Encryption avoids the contents of the migrating agents and messages to be picked up.

2.2 Functionalities of Agents

When the platform receives messages which includes requirements described in the language mentioned later and are addressed to a field in the node from the user or other agents, the platform generates an agent in the field by the agent management facility.

After that, the agent behaves to satisfy the requirements by the following facilities.

Planning Facility

This facility generates action sequences to satisfy the requirements and execute the sequences. For example, the agents obtain knowledge about the state of the system as initial states and knowledge about the agents' actions from the knowledge manager and generates the actions sequences from the knowledge and execute them.

However, the following controls are necessary to deal with unexpected changes in open networks.

• Because there may occur inconsistency among knowledge about actions, the agents should interleave plan generation and execution and make sure the consistency of the knowledge.

• If knowledge often changes, it is better to obtain the knowledge just when it is needed. For this purpose, agents should do planning and knowledge gathering mutually. Moreover, those knowledge should be distinguished as uncertain knowledge.

• The agent mobility and inter-agent cooperation facilities described later are needed because the agents should gather information in a real-time way from the networks and act on remote nodes.

• Agents should use the blackboards in order to handle unexpected changes as quick as possible. When the agents' behaviors become inconsistent because of the unexpected changes, the agents write the information on the blackboards. Then related agents notice the information and they can modify their behaviors.

Mobility in Networks

As explained above, the agent mobility and inter-agent cooperation facilities are needed because the agents should gather information in a real-time way from the networks and act on remote nodes.

Open networks are often partly connected by lines with low reliability such as wireless connection for mobile computers. It is very inefficient to gather information and communicate through these lines. This problem can be solved by sending agents which do the job through the lines and make them behave in the remote nodes.

For example, in the planning phase, if agents need knowledge in remote nodes to make plans including actions in the remote nodes, the agents move to fields in the remote node, ask the knowledge managers of the fields and continue to make plans.

Inter-Agent Cooperation

When agents gather information from or act on remote nodes through reliable networks lines with broad bandwidth, it is desirable for the agents to cooperate with agents in the remote nodes by the following reason.

• Since agents in more than one node can act simultaneously, efficient processing is available.

• If there is inconsistency among processing in nodes, the agents can handle the inconsistency rapidly by sending information from one node to another.

The following two methods can realize inter-agent cooperation.

• Cooperation by message exchange between agents. The contract net protocol is one example.

• Cooperation by agents' access to shared blackboards. Layered blackboards can realize more flexibility.

In this proposal, it is proposed to be possible to deal with unexpected changes in open networks flexibly by combining these two methods.

Knowledge Management Protocol

Again in order to deal with unexpected changes in the open network, agents need to update knowledge managed by knowledge managers. For example, if the agents fail to make plans using knowledge obtained from the knowledge managers, the knowledge is turned out not to be correct. In this case, the knowledge should be updated.

For this purpose, the agents can ask the knowledge managers in the nodes to which the agents belong or in the remote nodes by moving to the nodes. The agents can also update the knowledge in the remote nodes by cooperating with the agents in the nodes.

2.3 Description Language

As described above, it is required that agents have various complex facilities such as planning, mobility and cooperation. To make efficient and easy use of these facilities, it is important to describe users' requirements and knowledge for planning in a special language. The language should have the following features.

• The language is based on a multi-agent model. That is, the language should support primitives such as message exchange, blackboards and basic cooperation mechanism, for example, the contract net protocol.

• It has primitives to manage networks explicitly. These primitives can handle ID's of nodes and fields.

• It has the "go" operator for agent mobility.

• It has primitives for planning and plan execution. Especially, there are primitives to invoke external commands.

• Descriptions in this language consists of agent definitions and field definitions. The agent definitions are composed by an agent ID, the ID of the field to which the agent belongs, and behaviour descriptions. The field definitions are composed by an field ID and field knowledge.

2.4 Interfaces

In this proposal, interfaces between components of the system are important to operate the architecture in the whole network. In this part, the details of the interfaces are explained.

Inter-Agent Interfaces

An interface for message exchange between agents is defined to realize inter-agent cooperation such as the contract net protocol. This interface consists of the following features.

• Primitives representing classifications of messages such as task announce or bidding in the contract net protocol. For example, the following primitives are represented by performatives of KQML.
  • The task announce is represented by the performatives "broadcast" and "achieve".

  • The bidding message is represented by the performative "reply".

  • The message for successful bidding is represented by the performative "broadcast".

• It is necessary to realize asynchronous broadcast on task announces and synchronous and asynchronous message sending addressed to definite agents as timing of message exchange.

Interface between Agents and Knowledge Managers

The information to be exchanged between agents and knowledge managers is as follows.

• Queries from agents to knowledge managers.

• Knowledge update inquiries from agents to knowledge managers.

• Replies from knowledge managers for the above messages.

Interface between Agents and Blackboards

This consists of the following features.

• Agents' access to blackboards such as read/write, deletion and update.

• Event-driven control of agents via blackboards.

User Interface

The following functionalities are required to the user interface.

• System setup on each node.

  It should be enabled to setup parameters and to manage fields such as generation, deletion and update.

• Knowledge management by users on each node.

  This includes generation of knowledge bases, knowledge input and update.

• Input of requirements by users

  Users should be able to input their requirements not only by the language described above but also by special visual formats proper to each application domain.

• Monitoring of agents

  The system should be able to display states of agents. It also should allow users to input commands to control agents.

  The states of agents consists of the following elements.

  • The condition of the agents' behaviors in the local node such as acting or pausing.

  • The condition of the agents in remote nodes.

3. Conclusion

An agent technique to deal with unexpected changes in open networks was proposed. In this proposal, inter-agent cooperation techniques and agent mobility are connected on the basis of planning techniques.

In order to improve the agent architecture in this proposal, the following techniques are necessary. However, it is still difficult to incorporate these techniques. Therefore they should be future work.

• Machine learning techniques make it possible for agents to correct knowledge automatically according to situations.

• For ensuring planning, it is necessary to unify the vocabularies (called ontologies) for knowledge representation among the nodes. This is temporally possible by restricting application domains, but frameworks for general purpose will be also possible using a technique called knowledge sharing.