Many challenges in today's society can be tackled by distributed open systems. This is particularly true for domains that are commonly perceived under the umbrella of smart cities, such as intelligent transportation, smart energy grids, or participative governance. When designing computer applications for these domains, it is necessary to account for the fact that the elements of such systems, often called software agents, are usually made by different designers and act on behalf of particular

more »

... takeholders. Furthermore, it is unknown at design time when such agents will enter or leave the system, and what interests new agents will represent. To instil coordination in such systems is particularly demanding, as usually only part of them can be directly controlled at runtime. Agreement technologies refer to a sandbox of tools and mechanisms for the development of such open multiagent systems, which are based on the notion of agreement. In this paper, we argue that agreement technologies are a suitable means for achieving coordination in smart city domains, and back our claim through examples of several real-world applications. Appl. Sci. 2018, 8, 816 2 of 38 the problem at hand [3]. However, many recent applications of multiagent systems refer to domains where agents, possibly built by different designers and representing different interests, may join and leave the system at a pace that is unknown at design time. It is apparent that coordination in such open multiagent systems requires a different, extended stance on coordination [3]. Application areas that fall under the umbrella of smart cities have recently gained momentum [4] . Intelligent transportation systems, smart energy grids, or participative governance are just some examples of domains where an improved efficiency of the use of shared urban resources (both physical and informational) can lead to a better quality of life for the citizens. It thus seems evident that new applications in the context of smart cities have the potential for achieving significant socioeconomic impact. We believe that applying AT to the domain of smart cities may enable the development of novel applications, both with regard to functionality for stakeholders, as well as with respect to the level of sustainability of smart city services. In particular, in this article, we discuss how coordination can be achieved in practical applications of multiagent systems, with different levels of openness, by making use of techniques from the sandbox of AT. Section 2 briefly introduces the fields of AT, coordination models, and smart cities, and relates them to each other. Section 3 describes several real-world applications, related to the field of smart cities, that illustrate how coordination models can be tailored to each particular case and its degree of openness. Section 4 summarises the lessons learnt from this enterprise. Background In this section, we introduce the fields of agreement technologies and coordination models and relate them to each other. We then briefly characterise the field of smart cities and argue that agreement technologies are a promising candidate to instil coordination in smart city applications. Agreement Technologies Agreement technologies (AT) [1] address next-generation open distributed systems, where interactions between software processes are based on the concept of agreement. AT-based systems are endowed with means to specify the "space" of agreements that can be reached, as well as interaction models for reaching agreement and monitoring agreement execution. In the context of AT, the elements of open distributed systems are usually conceived as software agents. There is still no consensus where to draw the border between programs or objects on the one hand and software agents on the other, but the latter are usually characterised by four key characteristics, namely, autonomy, social ability, responsiveness and proactiveness [5] . The interactions of a software agent with its environment (and with other agents) are guided by a reasonably complex program, capable of rather sophisticated activities such as reasoning, learning, or planning. Two main ingredients are essential for such multiagent systems based on AT: firstly, a normative model that defines the "rules of the game" that software agents and their interactions must comply with; and secondly, an interaction model where agreements are first established and then enacted. AT can then be conceived as a sandbox of methods, platforms, and tools to define, specify, and verify such systems. The basic elements of the AT sandbox are related to the challenges outlined by Sierra et al. for the domain of agreement computing [6], covering the fields of semantics, norms, organisations, argumentation and negotiation, as well as trust and reputation. Still, when dealing with open distributed systems made up of software agents, more sophisticated and computationally expensive models and mechanisms can be applied [7] . The key elements of the field of AT can be conceived of in a tower structure, where each level provides functionality to the levels above, as depicted in Figure 1 . Appl. Sci. 2018, 8, 816 3 of 38 Appl. Sci. 2018, 8, x FOR PEER REVIEW 3 of 36 Appl. Sci. 2018, 8, 816 4 of 38 of the designers in coordination mechanisms (micro-and/or macro-level properties), as well as different levels of control that designers have over the elements of the distributed intelligent system (the degree of openness of the system), as we will argue in the following [3]. Early work on coordination in multiagent systems (MAS) focused essentially on (cooperative) distributed problem solving. In this field, it is assumed that a system is constructed (usually from the scratch) out of several intelligent components, and that there is a single designer with full control over these agents. In particular, this implies that agents are benevolent (as instrumental local goals can be designed into them) and, by consequence, that the designer is capable of imposing whatever interaction patterns are deemed necessary to achieve efficient coordination within the system. Efficiency in this context usually refers to a trade-off between the system's resource consumption and the quality of the solution provided by the system: agents necessarily have only partial, and maybe even inconsistent views of the global state of the problem-solving process, so they need to exchange just enough information to be able to locally make good decisions (i.e., choices that are instrumental with respect to the overall system functionality). Resource consumption is not only measured in terms of computation but also of communication load. From a qualitative perspective, coordination in distributed problem-solving systems can be conceived as a distributed constraint problem (see [11] for an example). Agents locally determine individual actions that comply with the constraints (dependencies) that affect them, so as to give rise to "good" global solutions. Alternatively, in quantitative approaches, the structure of the coordination problem is hidden in the shape of a shared global multi-attribute utility function. An agent has control over only some of the function's attributes, and the global utility may increase/decrease in case there is a positive/negative dependency with an attribute governed by another agent, but these dependencies are hidden in the algorithm that computes the utility function and are thus not declaratively modelled. Quantitative approaches to coordination can be understood in terms a of a distributed optimisation problem. More recent research in the field of MAS has been shifting the focus towards open systems, where the assumption of a central designer with full control over the system components no longer holds. This raises interoperability problems that need to be addressed. In addition, the benevolence assumption of distributed problem-solving agents needs to be dropped: coordination mechanisms now have to deal with autonomous, self-interested behaviour-an aspect that is usually out of the scope of models from the field of distributed computing. Approaching agent design in open systems from a micro-level perspective means designing an intelligent software entity capable of successful autonomous action in potentially hostile (multiagent) environments. In this context, coordination can be defined as "a way of adapting to the environment" [12]: adjusting one's decisions and actions to the presence of other agents, assuming that they show some sort of rationality. If the scenario is modelled within a quantitative framework, we are still concerned with multi-attribute utility functions, where only some attributes are controlled by a particular agent, but now there are different utility functions for each agent. The most popular way of characterising a problem of these characteristics is through (nonconstant sum) games [13]. Coordination from a micro-level perspective boils down to agents applying some sort of "best response" strategy, and potentially leads to some notion of (Nash) equilibrium. From a macro-level perspective, coordination is about designing "rules of the game" such that, assuming that agents act rationally and comply with these rules, some desired properties or functionalities are instilled. In the field of game theory, this is termed mechanism design [13]. In practice, it implies designing potentially complex interaction protocols among the agents, which shape their "legal" action alternatives at each point in time, as well as institutions or infrastructures that make agents abide by the rules [9] . From this perspective, instilling coordination in an open multiagent system can be conceived as an act of governing interaction within the system. If the environment is such that agents can credibly commit to mutually binding agreements, coordinating with others comes down to negotiating the terms of such commitments. This is where the link to AT becomes evident. Norms and organisations define and structure the interactions that Appl. Sci. 2018, 8, 816 5 of 38