Making SharedPlans More Concise and Easier to Reason About Extended Abstract

When a group of agents get together to collaborate on some complex group action, collaboration does not just happen. It requires the existence or formation of mutual beliefs about the capabilities and commitments of agents responsible for doing various actions, the adoption by individual agents of various intentions (not only intentions-to do actions, but also intentions-that various propositions hold), and a process of plan “elaboration” whereby a partial plan is expanded toward completion as a full plan. Grosz and Kraus’ SharedPlans [1, 2] is a general theory of collaborative planning that accommodates multi-level action decomposition hierarchies, models the collaborative support provided by group members to those agents or subgroups responsible for doing constituent actions, specifies what it means for a group of agents to have a partial plan, and explicates the process whereby a partial plan may be elaborated into a full plan. In the original formulation of SharedPlans, henceforth called V1 (for Version 1), meta-predicates model the plans of individual agents and groups of agents. Because the meta-predicate definitions use recursion interleaved with existential quantification and modal belief operators, reasoning about V1 SharedPlans is difficult. Furthermore, much of the complex structure of a V1 SharedPlan is left implicit. This paper, and its longer version [3], present a reformulation of SharedPlans, henceforth called V2, that (1) introduces SharedPlan Trees to make explicit the complex structure of SharedPlans, (2) simplifies and reorganizes the SharedPlan meta-predicate definitions without sacrificing their expressiveness, and (3) enables conditions to be specified under which a set of important theorems about agents and their SharedPlans may be proven.