Research On The Learning Of Action Model For Temporal Planning

Temporal actions are models of how agents can change the world around them,in terms of logical relations,temporal relations and resource-change relations.Building such models by human experts from the ground requires a large amount of domain knowledge and careful investigations,and therefore is time consuming and laborious.In recent years,automating the process of action model construction,to aid human experts,has been paid more and more attention,and becomes a hot topic in the area of Automated Planning.This work aims at developing a set of methods to automatically learn the structure of temporal actions specified in PDDL2.1.In this work we not only extended existing methods for STRIPS action modelling but also designed new methods for some of the key problems for PDDL2.1 specified action models.As PDDL2.1 is more expressive than STRIPS,the work is helpful in applying automated action modeling in more complex environment.The learning of PDDL2.1 specified temporal actions were broken into the following phases: state partition(SP),logical relations learning(LRL),action duration learning(ADL),temporal relations learning(TRL).The results of these phases are then synthesizer into a complete temporal action model.Specifically,this work includes:(1)In the SP phase,we reviewed existing methods,and proposed a new method called "state partition by word vector"(SPWV)."Word vector" is a state of the art method in modelling words in the area of natural language processing(NLP).And in our evaluation,SPWV shows to be with lower error ratio when compared with a representative method.(2)In the ADL phase,we designed a two phased progression techniques,which resulted a method that is able to fit the arithmetical expressions of action duration with lower computational cost.(3)In the TRL phase,we designed a time-stamp analysis technique to learn the temporal relations among the action,their preconditions,and their effects.(4)The proposed methods were implemented in a system called “Temporal Action Modeler”(TAM).TAM was evaluated using benchmarks from the International Planning Competitions(IPCs).Results show that the proposed methods are of availability,and tend to obtain higher accuracy when they gain more observations.