Deliberative and BDI Agents

Two applications can be stated as deploying deliberative agents. The first is described in [WT92], in which an interesting agent framework is applied to control a VPN service.

Telecommunication services such as VPN rely on two separate layers: The logical layer maps the view of the customer’s network, which is built using the physical layer that maps the actual physical network of the operator. Here, agents are used to automate the negotiation that used to occur between the service provider and the service customers when customers ask to change the service parameters or to repair network faults. A cus-tomer agenthas knowledge about the logical structure of the VPN and its usage, while the provider agentknows both the logical and physical implementation of each customer’s VPN. Thus, agents have a model of the external world on which they act. Precisely, a cus-tomer agent knows for each trunk (i.e. logical link) its capacity and its utilization, whereas the provider agent knows for each logical trunk the physical links upon which it is con-figured.

When a network fault occurs, the VPNs based on the faulty network element are af-fected. In this case, the customer agent first asks the provider agent to repair the fault.

In many cases, a complete and immediate repair is not possible and a negotiation-based cooperation is started. Customer agents try to find intermediate solutions based on their knowledge of the utilization of their respective trunks, and suggest partial solutions to the provider agent by updating the logical structure of their VPNs. The provider agent coor-dinates these solutions and makes the necessary updates in order to reach an acceptable configuration. Both customer agents and provider agents use logical reasoning based on their models of the network and on their beliefs on its elements. Agents developed in this application make use of both, Shoham’s Agent Oriented Programming [Sho93] and planning functions provided on a lisp-like planner called PRODIGY [FV94].

The second application is carried out within the MANIA project (Managing Awareness in Networks with Intelligent Agents) [OL95]. During his PhD Thesis, Oliveira [Oli98]

developed a BDI agent-based approach to manage application-level quality of service in enterprise networks. Agents are structured in beliefs, desires, intentions, goals and commitments. A first part of the beliefs describe the real-time state of the network, e.g.

‘the printing service is being highly solicited at the present time’. A second part contains the historical behavior of the network. The historical behavior is used to achieve some kind of learning about the dynamics of the network such as deducing that the printing service is highly solicited every day from 10 to 11 am. A third part of the beliefs translate the states of the services provided over the network, e.g., the minimal response time the NFS server can ever have, or the maximum client number a web server can handle at a time. Finally, the agent may have beliefs on the user application contexts, i.e. user requirements in terms of QoS.

The agent desires consist of two parts. The first part corresponds to the requests that the agent could not satisfy, such as when a certain user requires a video connection while there is no available bandwidth (according to the agent belief). The second part consists of motivation policies. The network administrator may want to motivate the agent to give a certain priority to some project members because they have a constraining deadline.

When receiving a user-application context (which describes thre resources that the user needs and the required QoS parameters), the agent translates it into a set of goals.

Goals are independent from the system state. For example, a goal may state to ‘actively monitor the NFS server’. These abstract goals are then mapped into intentions. Inten-tions take into account the current state of the system, the available means to achieve the goals (e.g., MIBs, testers, etc.) and the possible constraints that may apply.

Finally, and in the same way that the administrator could motivate the agent, he is also able to specify obligation policies that form the commitment part of the agent.

3.4.3 Hybrid Agents

As a general rule, hybrid agents are able to exhibit both reflexive, therefore prompt be-havior, as well as deliberative, therefore long-term behavior. In general, such kind of agents are applied both, to perform local management functions within a network do-main, as well as to perform global management functions using their deliberative capa-bilities. Some agent-based network management approaches followed this principle.

An ATM network is very dynamic and managing it centrally will not be appropriate because managed data quickly go out of date. But local management is not appropriate because it lacks an overall view of the network. Here, the hybrid agents are particularly

interesting. They may combine both local real-time management via the reactive behav-ior, while the deliberative behavior is concerned with cooperating with other agents in order to achieve global planning and coordinate high-level tasks. HYBRID, proposed by Somers [Som96b, Som96a], is an agent-based framework for the overall management of ATM networks that makes use of hybrid KQML-speaking agents. HYBRID uses a hier-archical agent system in which authority is transferred from high-level agents to lower-level agents. From an agent architecture standpoint, HYBRID provides an inheritance hierarchy of agent types. The root agent is thebasic agent that only “supports KQML front-ends and conforms to some minimum behavior”. From this basic agent, several other agent types are inherited with increased capabilities. For example, asystem agent has “low-level skills” which are scripts called in response to events that occur in the man-aged system. Therefore, the system agent can be considered as reactive. Thetask agent is able to manage his concurrently running tasks. TheIntelligent agent which is in the bottom of the inheritance exhibits both reactive and deliberative behaviors. The reactive behavior is indirectly inherited from the system agent. The deliberative behavior is sup-ported by worldview knowledge bases, procedural skill bases, a situation assessor, and an agenda. The worldview knowledge bases include facts on the other agents, on the agent itself, and on the managed components. Procedural skill bases include low-level skills for the reactive behavior and high-level capabilities such as plans and negotiation skills.

The situation assessor determines the focus of attention of planning tasks according to the current agenda and the agent goals. Finally, planning can be used in combination with skill-matching in order to plan for agent actions.

Agents in HYBRID are designed according to their roles that help deciding whether a reactive agent is sufficient or deliberative capabilities are needed. The use of a common communication language ensures that agents with different levels of intelligence are able to coordinate their activities.

Another application of hybrid agents is achieved by Wagner and is based on the con-cept ofVivid agents. Vivid agents [Wag96] comprise both a reactive and a deliberative part. However, the reactive part of the agent is not hardwired within the agent. Instead, the deliberative part may dynamically change the reactive behavior.Reagentsare partic-ular vivid agents with no planning capabilities.

Reagents are applied to perform the distributed diagnosis of distributed systems [FMNS97]. The network is partitioned into physical domains, each domain having its own diagnostic agent. The agent has a detailed knowledge model of its domain and min-imal information about the neighboring domains, mainly the address of their respective agents. The multi-agent system applies a distributed version of the Model Based Diag-nosis (MBD). The principle is that when an agent detects a fault, e.g. a lost connection,

it starts by performing a local diagnosis of its domain. This is the deliberative behavior of the agent, since it runs MBD reasoning according to the model of its domain. If it finds no local fault, then it sends the resulting observations to the neighbor agent. The latter performs the same procedure. If it finds the faulty element(s), it reports it to the first agent. If the first agent receives the fault observation from another agent, then it must ensure that the other neighbor agent is informed by forwarding the report. This is a reactive behavior in which the agent does not perform any reasoning.

In [HB98a, HBL99], Hayzelden et al. describe a heterogeneous MAS for the dynamic management of Virtual Path Connections² in ATM networks. Instead of using hybrid agents combining the reactive and deliberative behavior, the proposed architecture, Tele-MACS (for Telecommunications Multi-Agent Control System), relies on a set of reac-tive agents under the control of distinct deliberareac-tive agents. Reacreac-tive agents are located in thecontrol plane layer. The control plane layer includes agents with detailed, but lo-cal knowledge and that operate at a fast time slo-cale. Deliberative agents are located in a management plane layer. The management plane layer includes agents with global ab-stract knowledge and that operate at a slow time scale. The whole architecture is inspired from the subsumption architecture, in which competences are organized in levels where the higher levels build new functionality upon the lower levels and have control on their function. This principle is applied in Tele-MACS in an elegant way. The approach is to let deliberative agents in the management plane layer provide the reactive agents in the control plane layer with tailored views of the world using the belief suppression mecha-nism [Hay99].