Plan coordination

Autonomous agents usually plan their actions. Sometimes agents can autonomously construct their plans and just benefit from cooperation, and sometimes they cannot even do without using resources from another. We distinguish two ways to facilitate cooperation by using each others resources. With service-based plan coordination agents offer resources they can produce for others. With request-based coordination agents are allowed to place requests for resources they need. These forms of coordination can be integrated in the plans, and be realized by adapting existing planning algorithms. All methods use a special model of the relevant state of the world and of possible actions. This model consists of predicates that are restricted to describe the properties of an available entity relevant for planning. Such a resource fact can be exchanged among plans more conveniently than sets of ordinary propositions as used by, e.g., STRIPS to describe a resource. Actions consume all resource facts that are given as a precondition, and produce the resource facts defined by the post-condition. One form of request-based coordination is when the plans of agents are coordinated after they have constructed their plans. This method, called plan merging is used for agents that are able to (first) create a valid plan on their own. Introduction When autonomous agents coordinate their actions, their combined potential increases significantly. An agent that plans its actions usually is far more efficient than a reactive agent. Similarly, we expect that agents that plan their interactions are even more efficient. We would like to have a method to plan the interactions among agents and combine this with the use of existing single-agent (refinement) planning methods (Kambhampati 1997). Moreover, in this approach, (i) agents should be able to decide themselves when they cooperate and which information they share, (ii) agents should be able to request services from other agents and include the results in their plans, and (iii) agents should be able to offer services to other agents and, upon a request, add these to their plans. Such a method is called a multi-agent planning method (Georgeff 1984; Mali & Kambhampati 1999). Many planning and coordination problems can be modeled as a multi-agent planning problem. For example, the coordination of transport organizations, armies, production processes, etc., to name just a few. Currently, quite a body of research is focused on such problems, but “most of this research is still in its early stages, and many research challenges remain” (DesJardins et al. 2000). In this paper we focus on several autonomously planning agents that can interact through the exchange of resources without needing to provide more information if they do not want to. An interaction between two plans occurs if the effect of one agent’s action is used by the other. Usually, the effect of an actions is described by a set of propositions. We propose to use a special predicate for each physical object that the agents can exchange. This predicate denotes that this object is available and describes all attributes of such an object. We call such a predicate a resource fact. First, we give a more formal definition of a resource fact and we define actions as processes that consume and produce these resource facts. Then, we describe two ways to plan the coordination of agents: service-based coordination where an agent can offer to produce certain resource facts, and a method where agents publish requests for specific resource facts, i.e., request-based coordination. Finally, we present a special form of request-based coordination, called plan merging, where agents first construct their plans and then request replacements for resource facts in their plans to be able to remove some actions. Resource facts, actions and plans The Action Resource Framework, abbreviated ARF (Tonino et al. 2002; de Weerdt et al. 2003), distinguishes two basic notions in planning: resource facts and actions. Goals and plans are derived notions that are defined using resource facts and actions. A resource fact describes an object that is relevant to an agent with respect to the planning problem at hand. Such a resource fact represents either a physical object such as a truck or a block, or an abstract conceptual notion such as the right to do something. While using ordinary (classical) propositions (STRIPS) we need a set of propositions to describe all properties of such an object, in the ARF, these properties are be combined in one resource fact. Syntactically, a resource fact is denoted by a predicate name together with a complete specification of all its attributes. The predicate name serves to indicate the type of resource fact (e.g., a carrier cycle or a taxi). To uniquely identify resource facts, a special attribute identity is used to distinguish it from other resource facts having the same type and possibly the same values of their attributes. Because of the special nature of this identifier, we denote a resource fact of type with identifier and attributes of sorts respectively, as . For example, we describe a truck at a location , at 10 o’clock, and with left capacity 8 by