A Domain-Independent System for Case-Based Task Decomposition without Domain Theories

@inproceedings{citeulike:687413, author = {Xu, Ke and Munoz-Avila, Hector }, booktitle = {Proceedings of the Twentieth National Conference on Artificial Intelligence (AAAI-05)}, citeulike-article-id = {687413}, comment = {The authors devised a system in which the methods, or domain theories, in an HTN may be learned from a set of operators and variable type ontology, as well as a set of cases -- solved problems. These cases may be represented as methods in which the head is the goal of the problem, the preconditions are the start state of the problem, and the subtasks are the operators used in the solution. It seems to me that these cases would only ever call non-primitive subtasks, and so it is not immediately clear what value the hierarchical nature provides. New cases are generalized by substituting typed variables for all of the constant symbols used in the case, and insisting that variables with different names actually be different. At first this confused me, but I now realize it is because different constant symbols are by definition different. Constant preferences are then added to the cases such that each case prefers to bind its variables to the same constants that they originally represented. Finally, type preferences are added to the cases such that cases with more specific types will be preferred over cases with general types of their variables. When a new problem is encountered, cases are rated by their applicability to the problem by their ability to complete the task and their preconditions being met, then by whether or not they came from the same problem (constant preferences are satisfied), then by their level of specificity (type preferences are satisfied). Ungeneralized cases are guaranteed to provide a correct plan for all previously seen problems, but will be unable to provide plans for new problems. Generalized cases may be able to find solutions to new problems, but they are not guaranteed to provide correct solutions to any problem. Adding constant and type preferences guarantee that at least previously seen problems will be solved correctly, which they call the soundness property. An experiment is reported in which a case-based planner attempts problems in two domains using the different styles of cases. Although there was great variation, using generalization and constant and type preferences provided the best performance, in terms of a balance between precision and recall. It is not clear exactly where the cases came from, but my guess is that they were generated by a traditional HTN planner with domain theories. These tests used a threshold \$\alpha\$ such that planning would only be attempted if the best case had similarity to the problem of at least \$\alpha\$. It is not clear why this is necessary, or how its seemingly arbitrary values were chosen.}, keywords = {cbr, htns, learning, planning}, posted-at = {2006-06-06 19:56:06}, priority = {0}, publisher = {AAAI Press}, title = {A Domain-Independent System for Case-Based Task Decomposition without Domain Theories}, year = {2005} }

TY - CONF ID - citeulike:687413 L3 - citeulike-article-id:687413 TI - A Domain-Independent System for Case-Based Task Decomposition without Domain Theories T2 - Proceedings of the Twentieth National Conference on Artificial Intelligence (AAAI-05) PB - AAAI Press KW - cbr KW - htns KW - learning KW - planning N1 - The authors devised a system in which the methods, or domain theories, in an HTN may be learned from a set of operators and variable type ontology, as well as a set of cases -- solved problems. These cases may be represented as methods in which the head is the goal of the problem, the preconditions are the start state of the problem, and the subtasks are the operators used in the solution. It seems to me that these cases would only ever call non-primitive subtasks, and so it is not immediately clear what value the hierarchical nature provides. New cases are generalized by substituting typed variables for all of the constant symbols used in the case, and insisting that variables with different names actually be different. At first this confused me, but I now realize it is because different constant symbols are by definition different. Constant preferences are then added to the cases such that each case prefers to bind its variables to the same constants that they originally represented. Finally, type preferences are added to the cases such that cases with more specific types will be preferred over cases with general types of their variables. When a new problem is encountered, cases are rated by their applicability to the problem by their ability to complete the task and their preconditions being met, then by whether or not they came from the same problem (constant preferences are satisfied), then by their level of specificity (type preferences are satisfied). Ungeneralized cases are guaranteed to provide a correct plan for all previously seen problems, but will be unable to provide plans for new problems. Generalized cases may be able to find solutions to new problems, but they are not guaranteed to provide correct solutions to any problem. Adding constant and type preferences guarantee that at least previously seen problems will be solved correctly, which they call the soundness property. An experiment is reported in which a case-based planner attempts problems in two domains using the different styles of cases. Although there was great variation, using generalization and constant and type preferences provided the best performance, in terms of a balance between precision and recall. It is not clear exactly where the cases came from, but my guess is that they were generated by a traditional HTN planner with domain theories. These tests used a threshold $\alpha$ such that planning would only be attempted if the best case had similarity to the problem of at least $\alpha$. It is not clear why this is necessary, or how its seemingly arbitrary values were chosen. AU - Xu, Ke AU - Munoz-Avila, Hector PY - 2005/// ER -