Novel approaches for generalized planning
- Autores: Damir Lotinac
- Directores de la Tesis: Anders Jonsson (dir. tes.)
- Lectura: En la Universitat Pompeu Fabra ( España ) en 2017
- Idioma: español
- Tribunal Calificador de la Tesis: Víctor Dalmau Lloret (presid.), Sergio Jiménez Celorrio (secret.), Marco Aiello (voc.)
- Programa de doctorado: Programa de Doctorado en Tecnologías de la Información y las Comunicaciones por la Universidad Pompeu Fabra
- Materias:
- Enlaces
- Tesis en acceso abierto en: TDX
- Resumen
Classical planning is the problem of finding a sequence of actions, which leads from an initial state to some goal state. The effects of the actions are deterministic and the states are fully observable. It is equivalent to a search problem in an implicitly defined directed graph, where states represent the nodes, and actions represent the edges. To tackle the complexity arising from the number of states, modern planners usually rely on heuristics. The heuristics are usually computed from a relaxed version of the original planning problem. However, such planners find solutions for a single instance of the problem. Each time a classical planner solves a planning instance, the search begins without any prior knowledge.
In generalized planning, a plan is a solution to a set of planning problems, which belong to the same class. In this thesis we explore novel ways of computing generalized plans, inductively from a set of examples, and deductively from a model of actions. First we present an extension of planning programs, as a representation of a generalized plan, which is induced from a set of examples. The extension, allows for modeling of classification tasks. This work also introduces a novel domain-independent algorithm for generating hierarchical task networks directly from the action model and one representative instance of the planning problem. We also present the optimizations used by the translation and show that the algorithm is competitive with the state-of-the-art algorithms.
The main contributions of this work are:
The extension of planning programs, which allows for usage of high-level state features in the form of conjunctive queries. Moreover the algorithm couples the generation of the features with the generation of the program.
The noise-free classifiers represented by an extension of planning programs. While not competitive with ML classifiers, the high-level state features with planning programs bring a novel type of domain and possible benchmark to classical planning.
The domain-independent algorithm for generating hierarchical task networks based on invariance analysis. In this work we present a compilation which captures the abstractions, which can effectively reduce the search space in many planning domains.
A sound framework for HTN planning. We show that the HTN generator algorithm is sound in all presented variants. The algorithm also provides a sound framework, as certain types of decomposition methods can be removed or introduced without affecting the soundness of the plan.
The optimized version of generated HTNs. The introduced optimizations are domain-independent. Some of the presented optimizations work by pruning the search space and have possible implications on completeness, while others are used in form of the heuristic information. We show that the optimizations improve coverage and reduce backtracking over the IPC domains, and make the translation competitive with state-of-the-art HTN learning algorithms.
