Automated planning is known to be computationally hard in the general case. Propositional planning is PSPACE-complete and first-order planning is undecidable. One method for analyzing the computational complexity of planning is to study restricted subsets of planning instances, with the aim of differentiating instances with varying complexity. We use this methodology for studying the computational complexity of planning.Finding new tractable (i.e. polynomial-time solvable) problems has been a particularly important goal for researchers in the area. The reason behind this is not only to differentiate between easy and hard planning instances, but also to use polynomial-time solvable instances in order to construct better heuristic functions and improve planners.We identify a new class of tractable cost-optimal planning instances by restricting the causal graph. We also study the computational complexity of oversubscription planning (such as the net-benefit problem) under various restrictions and reveal strong connections with classical planning. Inspired by this, we present a method for compiling oversubscription planning problems into the ordinary plan existence problem. We further study the parameterized complexity of cost-optimal and net-benefit planning under the same restrictions and show that the choice of numeric domain for the action costs has a great impact on the parameterized complexity.We finally consider the parameterized complexity of certain problems related to partial-order planning. In some applications, less restricted plans than total-order plans are needed. Therefore, a partial-order plan is being used instead. When dealing with partial-order plans, one important question is how to achieve optimal partial order plans, i.e. having the highest degree of freedom according to some notion of flexibility. We study several optimization problems for partial-order plans, such as finding a minimum deordering or reordering, and finding the minimum parallel execution length.The research presented in this thesis has been partially funded by the National Graduate School of Computer Science in Sweden (CUGS).iii
Populärvetenskaplig sammanfattningHuvudtemat i denna avhandling är studiet av beräkningskomplexitet för automatisk planering. På en hög nivå kan man betrakta planeringsproblemet så här; man utgår från en värld eller ett system, som ett obemannat fordon eller en autonom robot, och antar att det finns vissa handlingar som ändrar tillståndet i systemet. Man antar dessutom att det finns ett initial-och ett måltillstånd. Planering är problemet att hitta en sekvens av handlingar som tranformerar initialtillståndet till måltillståndet. En sådan sekvens av handlingar kallas en plan. En planerare är ett datorprogram som löser planeringsproblem, och planerare baseras ofta på sökning genom tillstånden i systemet. Vanligtvis vill man optimera sökningen och hitta en plan som är optimal med avseende på något kriterium. Vad man vill optimera beror på tillämpningen men det är ofta att hitta en kortaste eller b...