Integrating Knowledge Acquisition in Plan Selection

Highly dynamic application domains represents a huge challenge for intelligent agents. Planning in an evolving environment involves problems such as catching discrepancies between real world properties and the agent representation of the world, and discriminating between relevant and non-relevant information.

The aim of this paper is to describe a novel approach to reactive planning in integrated architectures, where the plan selection mechanism is strictly related to the knowledge acquisition/verification process. The plan selection model is not limited to static testing of pre-conditions, but provides an integrated mechanism for the dynamic acquisition by need of that information which is relevant to a plan execution.

This result is obtained by partitioning the plan search space according to the actions which the planner is able to devise a plan for. Plans for each action are built according to an object-oriented perspective, in that they are designed according to the properties of the object involved. Correspondingly, each action induces a different view of the object of the application domain, leading to a particular configuration of the object in terms of knowledge chunk composition. As a result, a plan can be thought as an object method with respect to a given object configuration. Moreover, plan choice can be re-interpreted as a process of object dynamic classification, where a more refined view of n object brings to a more specialised plan for the corresponding action. The key-point is that such a refinement is possibly obtained through dynamic knowledge acquisition performed on demand, thus leading to a highly reactive mechanism of plan selection.