Automated design and con guration of complex systems is well established as a research topic, speci cally for constraint programming. In order to encode \quality of design" rules, i.e., preferences about the nal outcomes, the encoding must be completed with optimization capabilities. Based on experience in the con guration of elevator systems we present a research agenda to develop methods and tools which should be able to leverage constraint solvers in order to compute designs for complex systems that can compete with the ones produced by humans. Besides elevator systems, we propose to evaluate the methodology and the tools on other case studies, e.g., hardware con guration or heatingventilation and air conditioning systems.
In this paper we present and discuss the intention to build a tool able to handle
both con guration and design, e.g, in the terms of \quality" of the con guration
satisfaction [
In Section 2, after a short context de nition, we focus on elevators domain, which is our success story covered in Section 3. Encouraged by our results, we propose in Section 4 a detailed roadmap for the development of an automated con guration tool. Finally, in Section 5 we draw some conclusions on both the work carried out until now and the methodologies for future contribution. 2 2.1
Product con guration has been a fruitful topic of research in Arti cial
Intelligence (AI) since 1970, and also became a commercially successful application of
AI techniques. In the recent past, review and survey papers have been reported.
With focus on speci c topics on product con guration, these papers provide
rich information and knowledge in the corresponding areas, such as con
guration bene ts and challenges, con guration modeling and solving, con guration
knowledge handling and industrial cases, managerial issues related to con
guration in practice and theory testing of product con guration with respect to time
and quality performance. [
Con guration problems have been treated also as a class of constraint
satisfaction problems (CSPs) | see, e.g., in [
Satis ability Modulo Theories is the problem of deciding the satis ability
of a rst-order formula with respect to some decidable theory T . In particular,
SMT generalizes the boolean satis ability problem (SAT) by adding background
theories such as the theory of real numbers, the theory of integers, and the
3 GCSP is also de ned as Conditional CSP in [
Adding theories of cost to SMT yields Optimization Modulo Theories (OMT),
an extension that nds models to optimize given objectives through a
combination of SMT and optimization procedures [
Elevator systems are characterized by di erent lifting mechanisms, e.g., ropes and pistons, and di erent setups, e.g., one or two pistons, presence or absence of counterweights, and dedicated machine room versus in-plant machine room. In our work we focused on hydraulic elevators (HEs), which operate on hydraulics, i.e., they rely on one or more pistons to move the car. Energy is usually provided to the hydraulic uid by an electrically driven pump, and typically no counterweight is needed to compensate for the weight of the car. Their low initial costs, compact footprint and ease of installation makes them a viable choice for retro tting old residential buildings, and a cost-e ective solution for new ones alike. The choice of HEs as a case study is thus motivated by their popularity in low-rise applications, and by the fact that, in spite of their relative low part count, their structure presents already most of the challenges that are to be found in this domain. 3
3.1
4 Part of the AiLift suite available upon request at http://www.ailift.it. AiLift is a web-based application for the automated design of elevator systems. Those components must satisfy a set of constraints that guarantee the design compliance with respect to directive 2014/33/EU and related EN 81-20/81-50 norms. The problem of con guring this kind of system is that for both car frame and doors are selected from a database of existing components from di erent vendors and with di erent parameters.
Feeding a constraint solver with the domain outlined before yields a feasible con guration, if any. We may as well be interested in enumerating feasible conwcf Distance from xcf to the left car wall dcf External distance between car frame rails ygear Core gear placement with respect to the car frame base point dbr External depth of the car frame bracket from the base point dcr Distance between car rails maxoh Maximum car overhang that the car frame is able to sustain opening Doors opening lacd, racd Left and right axis size (car door) lald, rald Left and right axis size (landing door) stepcd Car door step stepld Landing door step dstep Distance between doors wframe Landing door external frame width gurations, or in pinpointing \optimal" con gurations, i.e., designs which have desirable properties according to some best principles and practices in elevator engineering. In Table 3 we show a quick summary of the combinations among feasibility (SAT) and optimality (OPT) targets with search for a single assignment (Search) and enumeration of con gurations (Enum). The two questions that we found most relevant for practical purposes are feasibility and optimality in a search context. Checking whether a feasible con guration exists is useful in all the cases in which the size of the shaft and/or a restricted availability of components makes the design hardly feasible. Finding an optimal con guration, assuming that several alternative con gurations exists, is about stepping from plain con guration problems to full- edged automated design. However, in order to achieve optimality, we need to instruct the solver not just about what cannot be done, but also about what should be done to obtain a \good" con guration.
We encoded this con guration problem with the SMT formalism, not only providing the feasibility constraints but also connecting the system with the components database; in this way, the solver has the duty to select those components (car frame and doors) whose parameters allow a feasible (possibly optimal) design. This encoding is embedded in a Java application which orchestrates at an higher level the components relationships | the whole problem is still bigger than this module | and provides database transactions and drawing services. 4
In order to extend our work towards system-independent design, we need to build a framework in which the con gurator is accompanied by another system providing the correct interpretation and translation of a generic model in a set of variables, constraints and cost function(s).
Ontology-based modeling is widely used in order to build consistent models
from a structured description of a system: Ontology Web Language (OWL) [
As soon as the system implementation will be adequately ready, an experimental analysis campaign will test the prototype development. The availability of a research prototype will enable such analysis, targeted to assess the quality of the proposed solution. We expect interactions between experimental part and implementation part in a sense that implemented techniques may be subject to modi cations and re nements based on the experimental ndings. From this prototype release on, another test campaign will accurately cover the quality and correctness of the application. The evaluation will be carried out considering computational times as well as evaluation of the computer generated con guration designs by human experts of the model tested, in order to also assess the quality and correctness of the generated solutions. In order to provide reliable benchmarks we will test the prototype on our well-known application of elevators, but we also identi ed more uncorrelated domains like hardware con
guration and heating/ventilation systems. Hardware con guration refers to the problem of choosing, connecting and verifying a set of components considering multiple constraints related to physical occupancy, power consumption, equivalent circuits and memory availability just to mention some, and the domain also allows the re-con guration of an already con gured system where some constraints change and some components can change their connections. Heating/ventilation systems, on the other hand, are likely to present few constraints related to the number of machines and their parameters but pose a more important weight on the optimization of, e.g., number of radiators, power consumption, overall cost.
At the end of experimental analysis we expect that the prototype will be able to produce feasible con gurations of a given system and will be comparable with other constraint expression solutions. The ideal conclusion of the proposed system is the implementation of an interactive interface, i.e., making possible to interactively add, remove or edit constraints to the model in analysis. With this feature the system would then let the user experience a complete and controlled design process in which not only hard, model-relative feasibility constraints are present but it is also possible to add custom re nements on the y. 5
In the context of our LiftCreate design automation tool, we presented a constraint-based encoding to solve con guration and design for hydraulic elevators, which proved successful and pushes us to continue the integration of such methodology until the | possibly | complete design process with constraintbased formulations.
This success story leads to our agenda toward a model-independent automated con gurator for both con guration and design targets, creating a framework in which a generic model expressed in an almost standard language can be interpreted, translated in the core formulation and processed in order to obtain a design for the requested cost function.