A Consequence of Incorporating Intentions in Means-End Planning

This paper explores the consequences of attempting to use the intentions of an agent to make planning decisions. In particular it shows that, a priori sets of preconditions are incompatible with using intentions to define the conditions for successful execution of an action. Instead it suggests replacing preconditions with situated intentional reasoning and robust failure mechanisms. 1 Intentions and Planning Researchers [1; 2; 10] have advocated the use of intentions in AI approaches to planning. They argue that there are at least two advantages to be gained: (1) intentions can constrain solutions to possible problems, and (2) they can "establish standards relevance," determining when an action is appropriate, when it can be said to succeed, and when it can be said to have failed. What we will argue here is that if intentions are to fulfill these roles in a principled way, the common use of preconditions in action representations must be replaced by situated reasoning. The argument makes use of notions of positive and negative intentions, action relevance, and action success, which we define in the remainder of this section. 1.1 Defining Intentions We start with an intuitive definition of intentions, referring the reader to [1] and [2] for a more formal account. Intuitively, a positive intention is a commitment by an agent to perform an action at a given *This research is supported by ARO grant DAAL03.89.C.0031 PRIME. time. As such, positive intentions differ from goals, which are usually taken to be (features of) states that an agent wants to bring about. An agent can have a positive intention to act so as to bring about a state, but his or her commitment is still taken to be toward the action and not the state. An agent can also have negative intentions towards actions that it is committed to not performing. For example, an agent might have a negative intention towards breaking things: as such, the agent is committed to not performing any action that it believes could cause something to break. Notice that negative intentions, like positive intentions, are still commitments to actions rather than commitments to states. 1.2 Defining Relevance and Success An action will be said to be relevant to achieving some positive intention if by performing that action, an agent could satisfy that intention. Thus, actions can be used to achieve any of their effects, as opposed to having a limited set of a priori goals that they can be used to achieve, as for example in [15]. With the relevance of an action defined in terms of positive intentions, it is natural to define an its success as any execution of the action that results in the relevant positive intentions of the agent being satisfied. That is, an action is successful if the intentions it was selected to fulfill are satisfied by its performance. An action will be said to have failed if the intentions of the agent are not satisfied. (This allows actions to fail due to an agent’s misconceptions about their effects, as well as due to being performed badly.) 39 From: AAAI Technical Report SS-93-03. Compilation copyright © 1993, AAAI (www.aaai.org). All rights reserved.