Learning Hierarchical Task Networks by Observation Export

@inproceedings{citeulike:921715, author = {Nejatl, Negin and Langley, Pat and Konik, Tolga}, booktitle = {Proceedings of the Twenty-Third International Conference on Machine Learning}, citeulike-article-id = {921715}, comment = {This paper describes a process for learning teleo-reactive logic programs from plan traces. Teleo-reactive logic programs are a special class of HTNs in which methods (known as non-primitive skills) must have precisely one effect. This is accomplished by the use of concepts, which are similar to axioms in standard HTNs. For a more detailed description of teleo-reactive logic programs, see Learning Recursive Control Programs from Problem Solving. The process of learning from traces is essentially the same as the learning process from problem solving, except that the entire structure is available from the beginning. Specifically, the learning process steps backwards through the operators applied, marking them as either skill chaining if the precondition of an operator is directly caused by a prior operator or true in the initial state, or as concept chaining if the precondition is a concept whose subconcepts are achieved by prior operators or are initially true. Non-primitive skills are learned as the algorithm drops out of each deeper recursive call by composing the current operator with previously completed skills. There are no existing systems that do a similar thing, but they compare to a macro-learning algorithm and an explanation-based method. The macro-learning algorithm performs so poorly that they do not report results. They compare the remaining systems by presenting them with a sequence of problems in which problems are directly solved if non-primitive skills to do so are known and otherwise the solution is analyzed to learn new non-primitive skills. The goal is to have as many problems fall into the first category as possible. Experiments on blocks world and depots (a version of logistics) demonstrate that the proposed method learns more (or more flexible) skills than the explanation-based method. The claim that hierarchical task networks are learned is true but disingenuous, as only certain classes of HTNs may be represented in this formalism. The paper provides a good overview of existing non-hierarchical learning from plan traces (as in macro operators).}, keywords = {htns, learning, planning}, posted-at = {2006-11-01 20:03:52}, priority = {0}, title = {Learning Hierarchical Task Networks by Observation}, year = {2006} }

TY - CONF ID - citeulike:921715 L3 - citeulike-article-id:921715 TI - Learning Hierarchical Task Networks by Observation T2 - Proceedings of the Twenty-Third International Conference on Machine Learning KW - htns KW - learning KW - planning N1 - This paper describes a process for learning teleo-reactive logic programs from plan traces. Teleo-reactive logic programs are a special class of HTNs in which methods (known as non-primitive skills) must have precisely one effect. This is accomplished by the use of concepts, which are similar to axioms in standard HTNs. For a more detailed description of teleo-reactive logic programs, see Learning Recursive Control Programs from Problem Solving. The process of learning from traces is essentially the same as the learning process from problem solving, except that the entire structure is available from the beginning. Specifically, the learning process steps backwards through the operators applied, marking them as either skill chaining if the precondition of an operator is directly caused by a prior operator or true in the initial state, or as concept chaining if the precondition is a concept whose subconcepts are achieved by prior operators or are initially true. Non-primitive skills are learned as the algorithm drops out of each deeper recursive call by composing the current operator with previously completed skills. There are no existing systems that do a similar thing, but they compare to a macro-learning algorithm and an explanation-based method. The macro-learning algorithm performs so poorly that they do not report results. They compare the remaining systems by presenting them with a sequence of problems in which problems are directly solved if non-primitive skills to do so are known and otherwise the solution is analyzed to learn new non-primitive skills. The goal is to have as many problems fall into the first category as possible. Experiments on blocks world and depots (a version of logistics) demonstrate that the proposed method learns more (or more flexible) skills than the explanation-based method. The claim that hierarchical task networks are learned is true but disingenuous, as only certain classes of HTNs may be represented in this formalism. The paper provides a good overview of existing non-hierarchical learning from plan traces (as in macro operators). AU - Nejatl, Negin AU - Langley, Pat AU - Konik, Tolga PY - 2006/// ER -