Planning to please: Following another agent's intended plan

Abstract

In this article we analyze a particular model of control among intelligent agents, that of nonabsolute control. Non-absolute control involves a "supervisor" agent that issues orders to a "subordinate" agent. An example might be a human agent on Earth directing the activities of a Mars-based semi-autonomous vehicle.Both agents operate with essentially the same goals. The subordinate agent, however, is assumed to have access to some information that the supervisor does not have. The agent is thus expected to exercise its judgment in following orders (i.e., following the true intent of the supervisor, to the best of its ability). After presenting our model, we discuss the planning problem: how would a subordinate agent choose among alternative plans? Our solutions focus on evaluating the distance between candidate plans, and is appropriate to any scenario in which one agent wants to follow (as much as possible) another agent's plan.

Searching for an alternative plan

“…Moreover, both these metrics have a flaw because they treat a plan as a set of states while a plan should more correctly be considered a sequence of states (we would like the beginning of plan A to be close to the beginning of plan B, and so on). The shifting distance as defined in [3] might map a state at the beginning of one plan to a state at the end of the other plan, only because they are very close to one another. Figure 2 illustrates such a scenario.…”

“…There are several sensible ways to define such a distance metric in this domain. We introduce two metrics that were also proposed in [3].…”

“…Ephrati and Rosenschein [3] considered a situation consisting of two agents, one of them the supervisor and the other the subordinate. The supervisor created a plan for the subordinate, but the latter realizes that the supervisor's intended plan cannot be followed.…”

“…However, in the time dimension state X of plan B should be mapped to state Z of plan A. In that sense, the best distance metric suggested by [3] was the dynamic deviation metric, which sums the sequential relative deviation produced by operators of each plan, and is defined as…”

“…Suppose that an intelligent agent accepts as input a complete plan from a supervisor agent, human or automated (this is related to the "supervisor -subordinate" model from [3]). Provision of the complete plan implies that for the supervisor, how the goal is achieved is of some concern.…”