Publication: Generación de políticas para planificación heurística mediante metaclasificadores
Loading...
Identifiers
Publication date
2013-06-12
Defense date
2013-06-26
Authors
Tutors
Journal Title
Journal ISSN
Volume Title
Publisher
Impact
Export
Abstract
La planificación automática es una de las ramas de la inteligencia artificial, que tiene por finalidad la generación de planes de manera automática.
Para ello, se usan a menudo funciones heurísticas que guían el proceso de
decisión de qué operador escoger para transitar de un estado a otro, con la
finalidad de alcanzar las metas propuestas con el menor coste posible.
Estudios anteriores han demostrado que la inclusión del aprendizaje automático como soporte a la decisión del planificador, genera buenos resultados.
Así, mediante ejemplos de aprendizaje se conseguía generar un árbol de
decisión, que era suministrado al planificador.
Este proyecto de investigación tiene como finalidad principal evaluar si
esta tendencia positiva se mantiene si, en vez de usar un único árbol, se emplean
combinaciones de ellos al mismo tiempo. Mediante técnicas de bagging,
se consigue dividir el conjunto de entrenamiento inicial, para crear varios árboles
de decisión que recibe el planificador. Éste, a su vez, emplea distintos
algoritmos de combinación de la información aportada por los árboles.
El estudio experimental ha demostrado varios aspectos relevantes. En
primer lugar, el uso de estas técnicas ha supuesto una mejora respecto al uso
de un sólo árbol, resolviendo en ocasiones más problemas con menor coste.
Pero esto no es así en todos los dominios evaluados, existiendo casos para
los cuales no existen diferencias significativas en los planes generados. En
segundo lugar, se ha de considerar el hecho de que el uso de varios árboles
incrementa los tiempos tanto de entrenamiento como de validación, siendo
más notables en el primer caso. Lo mismo ocurre con la memoria (principal
y secundaria) requerida para el sistema, que se verá aumentada cuantos más
árboles se quieran emplear de forma simultánea. ______________
Automated planning is an artifi cial intelligence branch, whose objective is generating plans automatically. Heuristic functions are often used to help choosing which operator is selected to transit from an state to another, in order to reach the proposed goals with the lower cost. Previous studies demonstrated that the use of machine learning combined with planners generates good results. An example is the use of decission trees to guide the planner. The objective of this project is to evaluate if this positive trend is maintained using combinations of several decission trees at the same time, instead of using one single tree. Through the use of bagging techniques, the initial training set is split in order to create several decision trees which are received by the planner. This planner uses diff erent algorithms which combine the information provided by the trees. The experimental study has demonstrated some relevant aspects. First, the use of these techniques has involved an improvement over the use of a single tree, solving more problems with lower cost for some domains. Second, the use of several trees increases the training and testing time. The same thing happens with the memory (main and secondary) required by the system, which will be increased as more trees are used at the same time.
Automated planning is an artifi cial intelligence branch, whose objective is generating plans automatically. Heuristic functions are often used to help choosing which operator is selected to transit from an state to another, in order to reach the proposed goals with the lower cost. Previous studies demonstrated that the use of machine learning combined with planners generates good results. An example is the use of decission trees to guide the planner. The objective of this project is to evaluate if this positive trend is maintained using combinations of several decission trees at the same time, instead of using one single tree. Through the use of bagging techniques, the initial training set is split in order to create several decision trees which are received by the planner. This planner uses diff erent algorithms which combine the information provided by the trees. The experimental study has demonstrated some relevant aspects. First, the use of these techniques has involved an improvement over the use of a single tree, solving more problems with lower cost for some domains. Second, the use of several trees increases the training and testing time. The same thing happens with the memory (main and secondary) required by the system, which will be increased as more trees are used at the same time.
Description
Keywords
Inteligencia artificial, Planificación automática, Aprendizaje, Algoritmos, Planificación heurística