Executing Contingent Plans: Addressing Challenges in Deploying Artificial Agents

Abstract

The vast majority of research in automated planning focuses on generating a plan from an initial problem specification; from the theoretical properties of this task to the implementation details required to do so efficiently. While such work is often motivated by practical applications, there is far less understanding of the issues associated with executing plans in online environments. In this work we focus on this understudied area, and the challenges / opportunities that arise when executing complex plans. Unlike many works in plan execution, we consider a form of contingent plans as the source for execution; their complexity stems from the sophisticated representation of the action effects used to model the uncertainty in the world. The key contribution of our work is a proposed executor that can reason using the sophisticated action effects, and we demonstrate the impact this can have empirically. In support of an effective executor, we also consider (1) the connection between the execution context and the planner’s view of the state of the world; and (2) the separation between the execution of an action (the aspect that affects the outside environment) and the realization of its effects (the aspect that captures what has actually changed).