Solving High-Level Planning Programs

In this work, we consider a middle ground between automated planning (Ghallab, Nau, and Traverso 2004; Nau 2007; Green 1969; Weld 1999) and agent-oriented highlevel programming (Shoham 1993; Lespérance et al. 1995; Levesque and Reiter 1998; Rao 1996). Specifically, we propose a framework for high-level programming of autonomous intelligent agents using pure declarative goals. Automated planning allows the specification of behavior in a declarative manner, thus providing an abstract, flexible, and powerful mechanism that caters for flexible behavior: any conceivable way of achieving the desired outcome may be constructed (from first-principle). On the other hand, agent-oriented programming accommodates useful “knowhow” domain knowledge encoding the typical operations of the domain of concern since agent systems generally “act as they go”. Interestingly, the advantages of each of the two approaches are the weaknesses of the other. Lookahead planning is intrinsically difficult computationally since plans are built from first-principle, and it is not tailored for longterm behavior in changing domains, where the actual behavior depends on contingencies. Similarly, agent-oriented approaches typically rely entirely on procedural knowledge that ought to be crafted at design time and upon which the ultimate behavior of the system shall depend entirely: no “new” plans can be generated. In this paper, we propose a novel account that mixes programming with planning, thus leveraging on the advantages of both approaches. In concrete, we assume a typical planning domain D describing the dynamics of the world and a so-called planning target program T to be realized in D. Target T is a basically a high-level program composed of goals, both achievement and maintenance ones. Technically, the target is a transition system in which states specify choice points and transitions specify a pair of maintenance and achievement goals to be chosen by the agent and realized next. At any point in time, the external world and the target agent are in some of their respective states, and the agent decides, autonomously, which goal to achieve next in order to be in its following desire state (e.g., be at the airport). At the same time, it specifies a goal that has to be