The management of networked computer systems is rapidly growing in complexity due to the advent of Internet and Intranets. The paradigm has shifted from management of single systems to the management of complex domains, where an increasing number of machines are interconnected. In this context the traditional approach to performing configuration tasks, which consists of writing procedural applications, becomes inadequate. Planning techniques allow to dynamically synthesize action plans by using basic building blocks, coded with pre and post conditions. This approach is much more flexible in fact it’s possible to manage unexpected situations and updating of the system without changing the set of basic actions. As far as software for system management is concerned planning techniques almost represent a novel approach. In fact past efforts that we are aware of, though providing remarkable results, only tackled a limited subset of the problem. In order to solve complex configuration problems we need to build a powerful planner which intertwines the creation of partial plans (each corresponding to the achievement of a subgoal) with the resolution of threats among them (deriving from the interference among subgoals). The aim of this paper is to show the power and the limits of the application of an already existing planner to configuration tasks so as to understand which features are needed by a planner suitable for computer systems. We chose UCPOP [7], a regressive nonlinear planner. As a case study, we have applied UCPOP to a typical configuration problem: an example of disk management on an HPUX system [8]. It consists of a non trivial goal which requires a multil evel plan of basic action. We need to fill t he gap between the classical assumptions of a planning algorithm and the handling of a real system. Future work will concern the study of how to overcome these limits and the investigation on how to use Constraint Satisfaction techniques [6] in order to increase the eff iciency of the

We propose a planning architecture where the planner and the executor interact with each other in order to face dynamic changes of the application domain. According to the deferred planning strategy proposed in [14], a plan schema is produced off-line by a generative constraint based planner and refined at execution time by retrieving up-to-date information when that available is no longer valid. In this setting, both planning and execution can be seen as search processes in the space of partial plans. We exploit the Interactive Constraint Satisfaction framework [12] which represents an extension of the Constraint Satisfaction paradigm for dealing with incomplete knowledge. Given the uncertainty of the plan execution in dynamic environments, a backup and recovery mechanism is necessary in order to allow backtracking at execution time.

In this paper, we present an architecture for planning and executing in complex and dynamic environments. Such application domains are characterized by large amount of data, too wide and dynamic to be completely stored and maintained up-to-date. Planning and execution activities should deal with unpredictable action effects, unknown knowledge and dynamic information. We propose to use a planner and an executor based on the Interactive Constraint Satisfaction framework combining both search efficiency and data acquisition ability. Both planning and execution can be seen as search processes where constraint propagation, data acquisition and search are interleaved.

The authors' intrusion detection system integrates PlanNet, a constraint-based planner, to perform security management in computer networks. The system dynamically computes reconfiguration and repair plans, retrieving only the information needed to compute the plan.

Constraint Satisfaction techniques have been recognized to be effective tools for increasing the efficiency of least commitment planners. We focus on least commitment on variable binding. A constraint based approach for this issue has been previously proposed by Yang and Chan [21]. In this setting, the planning problem is mapped onto a Constraint Satisfaction Problem. Its variables represent domain objects and are defined on a finite domain of values; constraints remove inconsistent values from variable domains through constraint propagation. In many applications, however, it is not always convenient, if possible at all, to know in advance all objects belonging to variable domains. Thus, domain values should be retrieved during the plan construction only when needed. The interesting point is that data acquisition for each variable can be guided by the constraint (or the constraints) imposed on the variable itself, in order to retrieve only consistent values. For this purpose, we have extended a Partial Order Planner performing least commitment on variable binding. This extension can cope with incomplete knowledge. We use the Interactive Constraint Satisfaction framework defined in [12] in order to exploit the efficiency deriving from constraint propagation and the possibility of acquiring the domain knowledge during the plan construction. Experimental results and comparisons with related approaches show the effectiveness of the proposed technique.

Traditional planners build goal-oriented plans of actions by reasoning on a representation of the handled objects; this representation is assumed to be complete and static. On the other hand, planning needs to deal with the problem of incomplete and dynamic information when applied to real environments. The fields of network and security management are typical examples where it is really unrealistic making the assumption of complete and static world. In this paper, we present an intelligent support system performing planning tasks in the security management field. In such a context, a mechanism for dynamically acquire knowledge during the computational process is needed. Our approach is based on a definition of the planning problem as a Constraint Satisfaction Problem (CSP). We extended the CSP paradigm in order to deal with incomplete knowledge. Traditional CSPs work exclusively with completely known variable domains, while, in our solution, some variables can range on partially or completely unknown domains. The acquisition of domain values is performed by means of Interactive Constraints (IC), during their propagation. It is worth noting that only consistent information for the planner is retrieved so as to simplify further propagation steps.