Terminal de consulta web

Abstract: A planner must be prepared to deal with dynamic characteristics of the system it acts upon. Lack of world knowledge, dynamic changes, actions with non-deterministic results and faults may put the environment in a state different from the one the planner was expecting while performing a task. When this happens, the pre-conditions required for the plan may become false, turning the plan itself invalid and unable to reach the task goals. The problem of recovering from an invalid plan and achieving the task's original goals is called plan repair. Usually, plan repair techniques can be classified as either replanning techniques or conditional/probabilistic planning. Replanning consists of creating a new plan from the point of failure. Replanners are usually robust but too complex for using them in real-time applications. Probabilistic/conditional planners try to generate offline a plan that take into account the system's uncertainties. However, they might be unable to do so if the environment is too complex. In this work we propose a new plan repair technique based on action replacement. Any complex application will usually offer many different ways to achieve the results of any given action, although usually only the most efficient one is taken into account by the planner. In our replacement system, the planner will study the available devices and their relationship to build a replacement table, which lists subplans that can be used to replace a regular action with lessened efficiency (gracious degradation). We analyze the characteristics of the relationship between an action and its replacement subplans, and how to assemble the required table from this information. We describe the algorithm which implements the technique in a general context. We utilize this definition to apply the technique for simulated robots, a biped robot and web services