Automatic Induction of Bellman-Error Features for Probabilistic Planning

Every goal-oriented domain with a problem generator from the first or second IPPC (Younes, Littman, Weissman, & Asmuth, 2005; Bonet & Givan, 2006) was considered for inclusion in our experiments. For inclusion, we require a planning domain with a fixed action definition, as defined in Section 2.4 in the paper, that in addition has only ground conjunctive goal regions. Four domains have these properties directly, and we have adapted three more of the domains to have these properties as we describe in the next paragraph. The resulting selection provides seven IPPC planning domains for our empirical study. Figure 4 lists the reasons for the exclusion of the other six goal-oriented domains. In addition, four of the domains that we use in evaluation occur in both competitions in slightly different forms and we evaluate on one version of each of these four, as described in Figure 5. The three domains we adapted for inclusion are as follows. We created our own problem generators for the first IPPC domains TOWERS OF HANOI and FILEWORLD, as none were provided in the competition. For both these domains, there is only one instance of each size. In Towers of Hanoi, all instances share the same action set and state predicates, so that a suitable problem generator is straightforward. In Fileworld, a planning domain with a fixed action definition results if we consider the collection of instances that share the same fixed number of folders, but varying the number of files. When the number of folders varies, the state predicates and actions change, so that instances with varying numbers of folders cannot be in the same fixed-action-definition planning do-