An Extensible Action Architecture for Planning in Complex Domains

Planning in complex domains, including temporal and resource planning, often requires an expressive and flexible model of actions. For example, in a rover domain, the energy of the rover, a bottleneck resource, is consumed at a high rate while performing a drive action and in proportion to the distance traversed. Based on the current energy levels, the planning system can adapt the distance traversed due to the drive actions in order to best achieve its goals, i.e., if more energy is left, drive actions can be prolonged. To handle such domains, we consider an expressive model of actions containing variables (to denote, e.g., the energy consumed and the distance traversed within a drive action), and relations between the variables (e.g., the energy consumed is proportional to the distance traversed). We call these relations Action Component Relations (ACRs). ACRs may also be relations involving structures such as the conditions, effects, actuator tasks, etc., of the action, and other complex values (e.g., an ACR containing a set of temporal relations between the actuator tasks in a job-shop-scheduling domain). For efficient search, ACRs are optionally accompanied with pluggable domain-specific heuristics. In our extensible architecture, a programmer can easily add new kinds of ACRs. Finally, we present the use of ACRs in a story-planning domain for computer games. While our action specification model is generally applicable, our planning architecture is specific to utilize local-search-based planning.