Generalizing and Executing Plans

Abstract

Our work addresses the problem of generalizing a plan and representing it for efficient execution. A key area of automated planning is the study of how to generate a plan for an agent to execute. The plan itself may take on many forms: a sequence of actions, a partial ordering over a set of actions, or a procedure-like description of what the agent should do. Once a plan is found, the question remains as to how the agent should execute the plan. For simple forms of representation (e.g., a sequence of actions), the answer to this question is straightforward. However, when the plan representation is more expressive (e.g., a GOLOG program [4]), or the agent is acting in an uncertain world, execution can be considerably more challenging. We focus on the problem of how to generalize various plan representations into a form that an agent can use for efficient and robust online execution.

Srivistava et al. propose a definition of a generalized plan as an algorithm that maps problem instances to a sequence of actions that solves the instance [7]. Our work fits nicely into this formalism, and in Section 3 we describe how a problem (i.e., a state of the world and goal) is mapped to a sequence of actions (i.e., what the agent should do).