Design patterns are widely used in various areas of computer science, the most notable example being software engineering. They have been introduced also for supporting the encoding of automated planning knowledge models, but with little success. In this paper, we argue about the merits of design patterns, as an example of the broader class of reusable abstractions, in the automated planning context; particularly, we aim at drawing attention to their potential usefulness for the explainability of domain-independent planning systems.
something of an ad-hoc process, where the skills of knowledge engineers signi
cantly in uence the quality of the resulting planning application [
Studies on Knowledge Engineering for Planning and Scheduling (KEPS) have
led to the creation of several tools and techniques to support the design of domain
knowledge structures, and the use of planners for real-world problems [
A di erent yet complementary approach relies on the notion of reusable
abstractions (see, for instance, [
Design patterns remove the mentioned issue of KEPS tools, particularly the need to learn to use a dedicated tool, and can actively support the widespread use of best practices in the eld. Notably, a rudimentary form of design patterns is implemented in the well-known online editor planning.domains,1 as the \Misc PDDL Generators" plugin. This plugin provides the PDDL code for representing some common structures, such as grids and maps. There is not, however, a description of the problem that such structures are aimed at solving, and there is no justi cation for the speci c encoding that is suggested.
Unfortunately, aside from the mentioned planning.domains example, design patterns are not widely used by the AI planning community. A reason for that may be the lack of a centralised repository for managing such design patterns, the lack of a commonly agreed way for encoding them, and the fact that the encoding of planning knowledge models is still seen as a sort of ad-hoc artisan process, whose shortcomings can be dealt with by using high-performance planning engines rather than a proper engineering process.
Here we would like to highlight the important role that reusable abstractions, and particularly design patterns can play for AI planning knowledge models, beside the support for knowledge encoding and acquisition.
An area of growing importance for the automated planning eld, as well as for
AI in general, is that of explainability. The idea being that an AI system should
be able to explain, to some extent, its behaviour to stakeholders. Focusing on
the narrower topic dubbed XAIP, as for EXplainable AI Planning, a number of
recent works considered the problem of de ning what \explainable" means for an
automated planning system, explaining and describing generated plans, bridging
the gap between machines and human stakeholders, and designing approaches
to explain the behaviour of planning systems [
As noted in the previous section, domain-independent planning systems are
heavily dependent on the provided knowledge models; this is also true when
explainability is concerned. The behaviour of a planning system can not transcend
the knowledge model, and the importance of the planning knowledge model has
been well-argued [
It should therefore come as no surprise that design patterns can play a signi cant role in the XAIP eld. Design patterns can provide a valuable and well-de ned mean for supporting explanations. This is because, when design patterns are implemented in a planning model, the design pattern description becomes a valuable source of additional knowledge, that is not encoded in the model. Further, there is not even the need for nding a way to \attach" such additional knowledge to the standard planning models. The management of additional knowledge is a potential issue for planning systems [24]. Instead, having a centralised repository for planning design patterns (something similar to what Source Making2 does for software engineering, for instance) would allow planning models to be encoded and circulated as they are right now, and would support explanation engines in looking for patterns and descriptions in a structured manner. In fact, the XAIP eld can provide new momentum to the use and management of design patterns for planning.
In a sense, the use of patterns becomes then an explicit way for providing explanations, that can of course be complemented by the line of work that de
vises techniques for identifying known substructures of planning models, and
link them to an expected behaviour (see for instance [
For exemplifying our vision for the use of design patterns for explainability,
let us consider the Mobile design pattern introduced by Simpson et al. [
The problem description, the use of appropriate naming convention, and the
optional components can allow to deal with the classes of explanations that
Vallati, McCluskey and Chrpa de ned as context-related and assumption-related
[
A context-related question that can be answered using knowledge that comes with the Mobile design pattern would be: Why can the vehicle not move from location X to location Y? The pattern description includes the answer to this question, by stating that the movement can be allowed between adjacent locations in a map. Similarly, taking into account the optional components, the use of the design pattern can help to answer questions like Why did the vehicle not move immediately at the start of the plan? Here the fact that a vehicle needs a driver to move can help in addressing this question. Assumptions-related questions can then relate to the maximum capacity of the vehicle, for instance, or to the fuel consumption of the vehicle.
Given the provided example, it should be clear that to exploit design patterns for supporting explainability, there is the need for structuring the way in which the corresponding problem description, optional components, etc. are encoded in a repository, and for dedicated approaches able to identify the most relevant knowledge with regards to a given query. However, the planning community can take inspiration from the work done in other research areas, for instance Knowledge Engineering and Knowledge Management, that have decades of experience in addressing similar challenges. 4
Reusable abstractions, mostly under the form of design patterns, have been introduced for supporting the knowledge encoding and acquisition processes. However, they did not have great success in the KEPS eld.
In this paper, we point to the fact that design patterns can play a signi cant role also for supporting the explainability of the behaviour of a planning system. The bene t for Knowledge engineers derived by using design patterns in planning models can therefore be twofold: (i) design patterns allow to deal with common problems by reusing best practices, and (ii) they directly support the explanaibility of the knowledge models and, indirectly, of the overall planning system.
Future work is envisaged in identifying suitable way for storing and managing design patterns for automated planning, with focus given to the way in which the problem they address is described in an explainable-friendly manner, and to the planning languages that are supported. A general repository of design patterns for the planning community, as it is currently done for benchmarks, could foster their use and give momentum to this important area of research. 24. Vallati, M., McCluskey, L., Chrpa, L.: Towards explanation-supportive knowledge engineering for planning. In: Proceedings of XAIP (2018)