Publication: Aprendizaje de conocimiento de control para planificación de tareas
Loading...
Identifiers
Publication date
2006
Defense date
2006-07-20
Authors
Advisors
Tutors
Journal Title
Journal ISSN
Volume Title
Publisher
Impact
Export
Abstract
Esta tesis doctoral se centra en dos disciplinas de la Inteligencia Artificial: la planificación de tareas y el aprendizaje automático. La planificación en IA desarrolla sistemas de computación que resuelvan problemas cuya solución consista de un conjunto de acciones, total o parcialmente ordenadas, denominado plan que, mediante su ejecución, consiguen pasar de una situación inicial a otra situación en que se hacen ciertas una serie de metas o fines perseguidos .Debido al enorme tamaño del espacio de búsqueda que hay que explorar, se considera un problema PSPACE-complete [Vil Ander,1991]. Normalmente es necesario definir algún tipo de conocimiento o heurıstica que permita obtener eficientemente el plan. Por otro lado, se dice de un programa de ordenador que aprende apartir de la experiencia E, con respecto a alguna clase de area T y una medida de rendimiento P, si su rendimiento en la tarea T, medido mediante P, mejora con la experiencia E[Mitchell,1997]
El objetivo de esta tesis es el desarrollo de sistemas de aprendizaje automático aplicables a planificación, para extraer el conocimiento recontrol (o heurısticas) adecuado que incremente la eficiencia (disminuya el tiempo y la memoria consumidos por el ordenador) en resoluciones de nuevos problemas. También se busca definir una metodología de aprendizaje adaptable a distintos paradigmas de planificación capaz de obtener heurısticas. Por ultimo, se estudia la transferencia de dicho conocimiento, primero entre dos planificadores distintos; es decir, se extrae el conocimiento de una técnica de planificación específica, utilizando la metodología diseñada, para aplicarlo en otra diferente; y segundo, a los sistemas de aprendizaje :como le afecta al rendimiento del sistema de aprendizaje la utilización de un conocimiento previo proveniente de una fuente externa. Para poder realizar estas transferencias se necesita haber definido previamente un lenguaje de representación del conocimiento de control
__________________________________________________ The research presented in this dissertation focuses on two topics in Artificial Intelligence: planning and machine learning. The aim of classical planning is to find plans to solve problems. A problem is made of a starting situation and a set of goals. Given a set of possible actions, a planner must find a sequence of actions (a plan) that, starting from the initial state, fulfills all goals. Planning has a high computational complexity. In fact, it is PSPACE-complete [Bylander, 1991]. Usually, in order to make the search process more efficient, some heuristics are needed. On the other hand, a computer program is said to learn from experience E, with respect to some task T and a performance measure P, if its performance on T according to P, improves with experience E [Mitchell, 1997]. This thesis aims to show that machine learning methods can improve efficiency (time and memory) in modern planners, like graphplan-based planning or agentbased hierarchical partial-order planners. From this experience, a general methodology is developed to apply machine learning to different planning paradigms. Finally, this thesis explores the idea of transferring heuristics learned in one planner to a different planner. In particular, we transfer heuristics from a graphplan-based planner to a bidirectional one
__________________________________________________ The research presented in this dissertation focuses on two topics in Artificial Intelligence: planning and machine learning. The aim of classical planning is to find plans to solve problems. A problem is made of a starting situation and a set of goals. Given a set of possible actions, a planner must find a sequence of actions (a plan) that, starting from the initial state, fulfills all goals. Planning has a high computational complexity. In fact, it is PSPACE-complete [Bylander, 1991]. Usually, in order to make the search process more efficient, some heuristics are needed. On the other hand, a computer program is said to learn from experience E, with respect to some task T and a performance measure P, if its performance on T according to P, improves with experience E [Mitchell, 1997]. This thesis aims to show that machine learning methods can improve efficiency (time and memory) in modern planners, like graphplan-based planning or agentbased hierarchical partial-order planners. From this experience, a general methodology is developed to apply machine learning to different planning paradigms. Finally, this thesis explores the idea of transferring heuristics learned in one planner to a different planner. In particular, we transfer heuristics from a graphplan-based planner to a bidirectional one
Description
Keywords
Inteligencia artificial, Aprendizaje