- Abstract: Planning research has produced a large amount of tools for various formalisms. As planning is a computationally challenging task, it is important to come up with a variety of ideas and approaches to tackle the various sources of planning tasks complexity. However, as a by-product, it is unclear even to experienced planning researcher what tool will work well on a new planning task. The challenge is even harder for a layman. Most planning tools are not easily accessible and those that are might have non adequate performance on some tasks. The problem was partially addressed by the most recent International Planning Competition, with the competing planners being made publicly available in Singularity containers, allowing for easily building and running the planners. This, however, does not solve the problem of choosing the right planner for a given task. An impatient user might forgo the option of using domain-independent planners altogether as a result of an inadequate performance of one randomly chosen planner. Online algorithm selection using machine learning techniques was shown to be able to produce planners that show good performance on previously unseen domains. For optimal planning, previous success stories were mostly exploiting the fact that most tasks, if solved, are solved quickly. Thus, being able to accurately predict planner performance opens new perspectives for better exploiting various existing planners in practice. The research in this field does, however, currently require a rather deep familiarity with the field of planning. In order to allow researchers from outside of planning community to tackle the problem of planner performance prediction, we should alleviate the dependence on special knowledge in planning. One step towards achieving this goal is to provide data, complying with the assumptions made on data in the field of machine learning. The data consists of data points, some of which are labelled. Each data point represents a planning task, while labels represent planner performance on that task. The most important assumption about the data is independent and identical distribution (iid). Such an assumption is unrealistic when planning domains are created manually. Another assumption is that the data is representative of the entire population. In domain-independent planning, where the population consists of all tasks representable in the language of choice (e.g., PDDL), creating planning domains manually cannot produce representable data. For both, it is required to produce data automatically, and in a way that will cover a variety of possible planning tasks. We propose a new track at IPC to help achieving the mentioned goals. The track will provide an easy access to existing planners, as well as to the data -- a variety of hand-crafted, as well as automatically generated planning tasks, with an additional information that relates to the performance of these planners on the existing tasks. If providing this information seems to be unrealistic (e.g., due to computational load, as all planners would need to be run under the same conditions), an alternative could be to provide the instructions of how to obtain the performance information by running the planners (provided as containers as in the most recent IPC). Furthermore, instead of only providing a fixed set of benchmarks, one could also provide generators to automatically generate more data. In some sense, this new track would be similar to the learning track of previous IPCs, but with planners being provided alongside benchmarks. The goal of this new track would be to come up with a practical system that can perform well across planning domains. The uniqueness of this track should be in alleviating the need for special knowledge in planning. The goal is to both achieve better exposure and to ease the use of planning tools outside of planning community.
- Track: discussion topic (5 minutes)