Hierarchical landmarks - a means to reduce search effort in hybrid planning

Artificial Intelligence planning is a key problem solving technology currently being used in a variety of applications including military campaigns, robot navigation, airplane scheduling, and human computer interaction. The generation of plans courses of actions to achieve desired goals or perform specific tasks is a costly process, however. Developing methods to systematically reduce the search effort and increase (the) performance of planning systems is thus a central concern. In recent years, a number of approaches have been proposed to run a preliminary analysis of the artificial intelligence planning domain. They aim to extract and exploit knowledge from the domain model and problem description in order to reduce the planning effort. In general, pre-processing approaches can be done either ”off-line” analyzing the domain model before having access to a problemor ”on-line” analyzing the domain model and planning problem description together. We have developed a novel pre-processing technique to extract knowledge from a hierarchically structured planning domain and a current planning problem description which is used to significantly improve planning performance. This pre-processing technique enables pruning all branches which can be proven to never lead to a solution by identifying tasks that are not achievable from a certain initial situation. The efficiency of planning systems depends on the kind of planning search strategy which the planner uses. We developed novel domain independent strategies relying on the knowledge that is generated by pre-processing in order to guide the hierarchical planning processes more effectively towards a solution of a given planning problem. The complexity in real-world applications has led artificial intelligence planning researchers to develop algorithms and systems that more closely match realistic planning environments, in which planning activity is often distributed, and plan generation can happen concurrently with plan execution. Finally, our pre-processing technique in the context of hierarchical planning approach is integrated with a multi-agent based planning approach to decompose the original planning problem into a set of sub-problems each of which can then be solved separately. Our integration approach presents two different techniques to split the planning problem into a set of sub-problems: Dependent which constructs a set of dependent sub-problems and Independent which produces a set of independent sub-problems. Our empirical evaluation shows that the pre-processing technique improves performance because the dead ends can be detected much earlier than without pruning and that our search strategies outperform many other possible strategies. In addition, the integration between the pre-processing technique and a multi-agent based planning approach dramatically reduce computation effort. In general, our empirical evaluation proves that our approach improves the efficiency of planning systems on the tested domains.