Improving a Planner's Performance through Online Heuristic Configuration of Domain Models

The separation of planner logic from domain knowledge supports the use of reformulation and configuration techniques, such as macro-actions and entanglements, which transform the model representation in order to improve a planner’s performance. One drawback of such an approach is that it may require a potentially expensive training phase. In this paper, we introduce heuristic approaches for the online configuration of planning domain models. The proposed heuristics consider different aspects of PDDL-encoded operators for reordering such operators in the domain model, relying on the assumption that the way in which operators are encoded carries useful information about their expected use. Heuristics for Domain Model Configurations Operator ordering in domain models has shown to have considerable impact on the planning process (Howe and Dahlman 2002). Recently, Vallati et al. (2015b) developed an approach that automatically configures domain models by re-ordering their elements (e.g. operators, predicates). This approach relies on an expensive training phase, that requires the availability of a large number of training instances which are representative of the testing ones, in order to effectively configure a domain model for improving the performance of a given domain-independent planner. Here we propose heuristics for ordering operators in PDDL models. The underlying idea is that the way in which operators are encoded in PDDL carries some knowledge about the expected use of the corresponding actions. Such knowledge can thus be exploited for providing a more suitable ordering of operators, for improving the performance of the planner that will be used for solving the given problem. As we consider the typical domain-independent scenario, where configuration should be performed online, the focus on operators provides a good trade-off between the additional overhead and the potential impact on performance. Formally, given a planning domain model M, and the corresponding set of operators O = (o1, ..., om), we propose heuristics that, by considering some aspects of the operators in O, provide as output an ordered list of operators Oh. Operators are then listed in the domain model accordingly. We introduce five heuristics for ordering operators, Copyright © 2017, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved. that consider the following aspects: EFF, The number of effects; PRE, The number of preconditions; RAT, The ratio between effects and preconditions; NEG, The number of negative effects; and PAR, The number of parameters. Considered aspects are quick to compute, and can provide intuition about the expected use of operators. For instance, the presence of a large number of negative effects imply that the corresponding actions are strongly affecting the world, and could therefore be an indication that they are rarely used. On the contrary, the presence of very few preconditions can be an indication of actions that are often used, as the required condition can be easily satisfied. The ratio between effects and preconditions can give some further insights by considering both aspects at the same time: a high ratio value points to actions that have many effects and few preconditions, so that can be used often; a low ratio value may denote some more problematic actions, that require many conditions to be satisfied and have a limited impact on the world –but such limited impact may be of critical importance for achieving goals. Finally, the number of parameters is an indicator of the expected number of grounded actions. Each heuristic has two possible instantiations: ordering operators according to decreasing or increasing values of the considered metric. Hereinafter, we will use numbers to refer to the ordering, and letters for identifying the heuristic. For instance, EFF2 indicates that operators are ordered increasingly according to the number of effects, i.e., the first listed operator is the one with the least number of effects. In our implementation, ties are broken following the relative order of operators in the original PDDL model. Experimental Analysis We selected 8 planners, based on their performance in the Agile track of IPC 2014 (Vallati et al. 2015a) and/or the use of different planning approaches. Experiments were performed on a quad-core 3.0 Ghz CPU, with 4GB of available RAM and 300 seconds cutoff time. We considered all the domain models used in the Agile track of IPC 2014, but Maintenance, Visitall and Openstack. Maintenance and Visitall have a model composed by only one operator, and proposed heuristics aim at ordering operators within domain models. Openstack has a different model per each problem, where elements of problem and domain models are mixed. This can add noise to the empirical evaluation of the effecProceedings of the Tenth International Symposium on Combinatorial Search (SoCS 2017)