Scalability in planning with limited resources

Several temporal planning systems (e.g. Allen & Koomen, Descartes , parcPLAN) divide the planning process into two distinct phases: action introduction and plan verification. This thesis develops a detailed model of these two phases and their “interface”, and shows how the model can be exploited to enhance the efficiency of plan generation. The main focus is on plan verification, and the context is the temporal planner parcPLAN. Plan verification in parcPLAN is formally defined as a constraint satisfaction problem. The definition serves to clarify the structural dependencies of the problem and to identify limitations of existing methods. The approach in parcPLAN is extended in two key aspects. First, resource limitations are considered earlier in the search, and second, temporal and resource reasoning are integrated tightly through the use of probes. The resulting system, parcPLAN2, offers significant performance gains over parcPLAN. An extensive empirical study is undertaken to assess the performance of the overall planning approach in parcPLAN2. To gain perspective, parcPLAN2 is compared with two state-of-the-art planners, viz. IPP and Blackbox. On the benchmark experiments, parcPLAN2 outperforms IPP and Blackbox by a significant margin. The results show that the burden in plan verification can be managed effectively, but only up to a certain point. For certain types of problem parcPLAN2 scales well, but for others its effectiveness is limited.