Generating Data In Planning: SAS+ Planning Tasks of a Given Causal Structure

Anonymous

Abstract:

The need for data in planning has long been established, by, e.g., machine learning based approaches. The existing data, however, is quite limited. There exists only a relatively small amount of hand-crafted planning domains, mostly introduced through International Planning Competitions. Further, this collection of domains is not necessarily diverse: many of these domains are some variants of the transportation problem.
In this work, we alleviate the shortage in existing planning tasks by automatically generating tasks of a particular causal structure. Given any graph G, we show how to create a SAS+ planning task with the causal graph isomorphic to G. We create a large collection of planning tasks by randomly generating graphs of various structural restrictions and creating planning tasks for these graphs, but also, more importantly, we provide the community with a tool that allows for on-demand generation of additional, possibly larger tasks. Our experimental evaluation ensures that the generated collection is interesting for the current state of affairs in classical cost-optimal planning, showing the performance of state-of-the-art symbolic search and heuristic search based planners.

TL;DR: We show how a planning task with a given causal graph can be constructed

Keywords: data generation, planning tasks generation, causal graph structure