25th International Workshop on Database and Expert Systems Applications

A Multi-Agent Coordination Framework for Smart

Building Energy Management
Thanos G. Stavropoulos1,2, Emmanouil S. Rigas2, Efstratios Kontopoulos3, Nick Bassiliades1,2, Ioannis Vlahavas1,2
1
School of Science and Technology, International Hellenic University, Greece
2
Department of Informatics, Aristotle University of Thessaloniki, 54124, Thessaloniki, Greece
{athstavr, erigas, nbassili, vlahavas}@csd.auth.gr
3
Information Technologies Institute, Centre of Research & Technology - Hellas, 57001, Thessaloniki, Greece
skontopo@iti.gr

Abstract—This paper presents a novel energy management WSDL1, to model and exploit universal, platform-indepentent
framework for multi-agent coordination in smart buildings. The APIs for the various devices that need to be used in such
framework builds on top of an existing Service-Oriented settings. State-of-the-art AmI paradigms include applications
middleware for Ambient Intelligence, which offers sensor and of limited scale and targeting domains such as user comfort,
actuator functions of wireless devices. The middleware also office and home automation or even agriculture and healthcare
provides a semantics infrastructure that assists in authoring (Ambient Assisted Living – AAL) [4].
agent policies for reducing energy consumption and maximizing
user comfort. Each agent within the framework is responsible for This paper introduces a novel agent coordination
monitoring the environmental context and controlling the framework for energy savings in buildings, based on an
electrical appliances of a specific room. However, the collective existing AmI infrastructure. More specifically, the existing
behavior of the multi-agent system is controlled by a Coordinator deployment at the International Hellenic University2 (IHU)
Agent that approves or rejects the allocation of building consists of various wireless sensor and actuator networks,
resources in time, aiming at more “long-term” goals that are out homogenized by a Web Service middleware. The semantic
of the reach and scope of the individual Room Agents. The infrastructure includes an ontology and Semantic Web Service
agents’ underlying logic is expressed via defeasible logics, a annotations, which aid domain experts in dynamically
formalism offering intuitive knowledge representation and authoring energy-saving policies. The latter are expressed via
advanced conflict resolution mechanisms. human-intuitive defeasible logics and are distributed to
Keywords—multi-agent systems; ambient intelligence; smart
individual Room Agents residing in each office or room. A
environments; web services; context-aware systems single Coordinator Agent manages all underlying agents to
achieve large-scale energy saving while ensuring comfort.
Defeasible logics support rule superiority relationships, which,
I. INTRODUCTION in turn, allow us to define three rule layers of ascending
With dwindling fossil fuels, and the increasingly negative importance: preference, maintenance and emergency. Each
impact of climate change on society, several countries have Room Agent is able to submit its preference rules to the
instigated plans on a national level to reduce carbon emissions Coordinator Agent, which fuses them with maintenance and
[1]. According to studies, the building sectors in the US emergency policies for optimal long-term building
represent more than 40% of primary energy consumption [2]. management.
Considering that the majority of this energy is produced from
non-renewable sources, it becomes clear that, in order for II. SMART ENVIRONMENT INTEGRATION
global CO2 emissions to be reduced, the energy consumption This section presents the middleware layer of the proposed
of buildings has to be reduced as well. In this vein, advanced architecture, along with descriptions for hardware integration.
Artificial Intelligence (AI) techniques, such as Ambient First, the nature of AmI applications is entailing a dynamic,
Intelligence (AmI), can be deployed, in order to achieve real-time environment and heterogeneous devices. The Service-
efficient and real time management of energy-consuming Oriented Architecture (SOA) has been proved to be extremely
activities in buildings. suitable for such environments and used extensively to provide
AmI is one of the prominent computing paradigms amongst the necessary abstractions for application development [5].
Ubiquitous Computing (UbiComp) and Pervasive Computing SOA provides a higher level of abstraction, agnostic of
(PerComp). According to these approaches, ambient and non- platform and communication protocol device programming.
intrusive devices are scattered around the environment, in order Additionally, it enhances extensibility, as new devices can be
to sense and act appropriately to context. Consequently, such integrated via the authoring of new services under the same
systems are also called Context-Aware and are mainly based on framework. All services comply with a Web-wide universally
Semantic Web [3] constructs, e.g. ontologies, to model sensed
context in a machine-interpretable way. So far, AmI systems
have been largely based on Web Service technologies, such as 1
Web Service Description Language: www.w3.org/TR/wsdl
2
http://www.ihu.edu.gr/
The Smart IHU project was funded by the Operational Program
Education and Lifelong Learning, OPS 200056 (International Hellenic
University, Thessaloniki, Greece)

1529-4188/14 $31.00 © 2014 IEEE 126

DOI 10.1109/DEXA.2014.39
established API, the W3C WSDL language, which resolves functionality. The services follow the WSDL/SOAP
syntactic heterogeneity by type-defining service operations. description and communication standards and have been
implemented using Java, JAX-WS7. In the context of this
This work builds upon the aWESoME Web Service paper, the most significant operations are:
middleware for AmI introduced in [6]. The middleware itself
entails three distinct layers, one for hardware integration, one • Smart Plug Service: GetStatus, GetPower, SwitchOn,
that implements and provides (syntactically expressed) services SwitchOff.
and a semantic description layer on top. Our motivation for
building a middleware from scratch is twofold: (a) to define • Sensor Board Service: GetTemperature, GetHumidity,
semantics more efficiently, and, (b) to integrate hardware GetLuminance.
modules that existing middleware paradigms do not. However, • Smart Clamper Service: GetPower.
whenever possible, open-source software has been properly
deployed to interface with some parts of the hardware. • Z-Wave Service: GetMotion, GetCO2.

A. Hardware Integration Layer C. Semantic Layer

This layer consists of the so-called driver modules, i.e. The semantic layer of the architecture consists of two
essentially plug-ins developed specifically for each device components: (a) the ontological infrastructure, and, (b) the
family of the proposed system that are briefly presented below. semantically annotated service endpoints. The ontological
infrastructure entails the BOnSAI (Smart Building Ontology for
Smart Plugs: Smart Plugs3 are both sensors and actuators, Ambient Intelligence) ontology [8], suitable for describing
which monitor an electrical appliance’s status, power concepts relevant to Smart Buildings, Ambient Intelligence,
consumption and enable switching it on or off. They form Services and Sensor Networks. It also extends existing
wireless encrypted ZigBee networks managed by a single ontologies, such as the OWL-S8 upper ontology for services, for
controller and are widely popular in both retail and research interoperability outside the borders of our system. Meanwhile,
[6], [7]. The Smart Plug Driver module, implemented in Java, the aWESoME web services are annotated following the
is invoked every time a ‘get’ or ‘set’ operation needs to be SAWSDL lightweight standard, enhancing service discovery.
performed over the Smart Plug network and supports In the context of this work some annotations are used during
bidirectional communication. the dynamic authoring of rules by experts. Namely,
Sensor Boards: Wireless networks of Sensor Boards were SensorOperation or ActuatorOperation annotations on WSDL
deployed to measure ambient environmental properties. Each operations can be used in rule conditions and rule effects,
chosen board4 entails a multi-sensing platform for temperature, respectively.
humidity and luminance and forms a mesh open ZigBee
network. Unlike Smart Plug communication, Sensor Board
communication is one-directional, as boards push notifications D. Application Layer
to the server Sensor Board Driver module.
The application layer of the architecture makes use of both
Smart Clampers: The wireless Smart Clamper bundle5 web service abstractions and semantic descriptions in order to
measures the total building and Data Center consumption. The provide high-level functionality to users or software agents.
Smart Clamper Driver is a Java module that collects clamper Human users can monitor and manage the infrastructure
data to be retrieved by the Smart Clamper Service. through a series of applications such as:
Z-Wave: This protocol unifies sensors and actuators from • A desktop administrative application to monitor all
different manufacturers at communication layer level. The sensors and manually manage electrical appliances in
selected devices in our setting detect motion in each room and real-time;
CO2 concentration level at four main indoor locations. Being
the largest and most diverse family of devices, the Z-Wave • A web application of aggregated consumption history,
Driver module incorporates the open source Open Z-Wave statistics, waste metrics and environmental data for
library6 to collect data from virtually any compatible device. visitors, students and staff to build energy-awareness;
• Intelligent agents who automatically manage the
infrastructure, such as the multi-agent framework
B. Web Service Layer discussed in the next section.
Building upon integration modules, the middleware offers a
designated Web Service for each device family. In turn, each III. MULTI-AGENT COORDINATION FRAMEWORK
service offers multiple operations to support underlying The proposed multi-agent framework integrates two types
of agents: (a) Room Agents (RAs), each of which are assigned
3
http://www.plugwise.com the energy management of a distinct room within the building,
4
http://www.prismaelectronics.eu
5 7
http://www.currentcost.com/ Java API for XML Web Services: https://jax-ws.java.net/
6 8
https://code.google.com/p/open-zwave/ OWL-S: http://www.w3.org/Submission/OWL-S/

127
and, (b) a Coordinator Agent (CA) that is responsible for
accepting or rejecting the allocation of building resources
proposed by RAs. In practice, each RA ignores the existence of
other RAs and is only aware of its own set of rules (forming a
policy), which it submits to the CA for approval or rejection.
The proposed application architecture is shown in Fig 1.
The agents’ underlying logic is expressed via defeasible
logics, a non-monotonic logics formalism that delivers
intuitive knowledge representation and advanced conflict
resolution mechanisms [9]. In defeasible logics there are three
distinct types of rules:
• Strict rules are denoted by A → p and are interpreted
in the typical sense: whenever the premises are
indisputable, then so is the conclusion. An example of
a strict rule is r1: summerTime → tempHigh (“the
temperature is high during the summer time”).
• Defeasible rules are denoted by A ⇒ p and, contrary to
strict rules, they can be defeated by contrary evidence.
Two defeater examples are r2: tempHigh ⇒
switchOnCooler (“the cooler is typically turned on
when the temperature is high”) and r3:
highConsumption ⇒ ¬switchOnCooler (“the cooler
should not be turned on whenever there is high energy
consumption”).
• Defeaters are denoted by A x p and do not actively
support conclusions, but can only prevent deriving
some of them. In other words, they are used to defeat
Fig 1. Framework architecture and Agents distribution at
respective defeasible conclusions, by producing
the IHU premises
evidence to the contrary. A defeater example is: r2':
comfort (e.g. the desired room temperature during the winter
isOnCooler x ¬switchOnCooler (“when the cooler is months) and optimizing energy consumption. Also, through a
on, there is no need to turn it on again”), which can respective policy, the CA handles the power management
defeat e.g. rule r2 mentioned previously. scheme of the building (e.g. deactivation of certain appliances
Additionally, the superiority relationship is used for during late night hours), as well as potential emergency
resolving conflicts among defeasible rules. For example, given situations (e.g. activation of emergency lights in case of fire).
the defeasible rules r2 and r3 above, no conclusive decision can Thus, within the framework, there are three distinct layers of
be made about whether the cooler should be turned on or not. rules with escalating priority: (a) preference rules, handled by
But, if the superiority relationship r3 > r2 is introduced, then r3 RAs, (b) maintenance rules, and, (c) emergency rules – the two
overrides r2 and we can eventually conclude that the cooler will latter categories are handled by the CA. The notion of priority
not be turned on. In this case rule r3 is called superior to r2 and here refers to priority among defeasible rules, as discussed
r2 inferior to r3. Note that the relation > is acyclic. previously. The following sample use case illustrates the
functionality of the proposed framework in more detail.
The advantages of applying defeasible instead of classical
logics are outlined as follows: A. Sample Use Case – Rule Agents
• Defeasible logics allow for reasoning with incomplete Suppose that the following example involves two rooms, ‘A1’
information; this is a critical trait in AmI settings, and ‘A2’, within the smart building and that two RAs, RAA1
where perfect knowledge of the environment is very and RAA2, are responsible for each room, respectively.
hard, if not impossible, to achieve; Suppose, also, that the employees working in these rooms
have defined the following preference rule sets, which have
• They introduce non-monotonicity, which leads to a
more intuitive type of reasoning, much closer to human been assigned to the RAs9:
reasoning, where the emergence of new information # room ‘A1’ preference rules (agent RAA1)
can lead to abandoning (i.e. defeating) previously
established conclusions and adopting new ones. 9
For authoring defeasible logic rule bases, one could download one
As mentioned previously, RAs have their own user-defined of the available rule editors, like, e.g. SPINdle’s defeasible logic
policies, each of which consisting of a set of rules expressing theory editor (http://spin.nicta.org.au/spindle/tools.html), or the much
user preferences. These policies aim at maximizing user more flexible S2DRREd (Syntactic-Semantic Defeasible Reasoning
Rule Editor) [10].

128
 pA1-01: time(X), X > 1600 → afternoon #strict rule pA2-02: time(X), X > 1700 → ¬dayTime
pA1-02: time(X), X < 1600 → ¬afternoon #strict rule pA2-11: dayTime ⇒ switchOff(A2, light)
pA1-11: afternoon ⇒ switchOn(light) #defeasible rule pA2-12: ¬dayTime ⇒ switchOn(A2, light)
pA1-12: ¬ afternoon ⇒ switchOff(light) #defeasible rule
Conflicts are generated between pairs of rules: (a) with
# room ‘A2’ preference rules (agent RAA2) opposite conclusions, and, (b) where one rule concludes that a
pA2-01: time(X), X < 1700 → dayTime #strict rule device should be turned on, while the other rule concludes that
pA2-02: time(X), X > 1700 → ¬dayTime #strict rule it should be turned off and vice-versa. The latter case is
pA2-11: dayTime ⇒ switchOff(light) #defeasible rule expressed via conflicting literals [11], which, however, are not
pA2-12: ¬dayTime ⇒ switchOn(light) #defeasible rule further discussed in this work for brevity reasons. Conflicts can
be resolved via appropriate superiority relationships among
As already mentioned, preference rules primarily handle user each pair of conflicting rules. In the cases of conflicts between
comfort. According to the above preferences, the employee- rules belonging to the same rule set (e.g. both rules are
user determines the time thresholds that designate – by preference rules), the superiority relationship has to be
personal standards – when the lights in each room should be explicitly added to the knowledge base by the user. On the
turned on or off. other hand, the cases of conflicts between rules belonging to
different rule sets is handled automatically by the system, via
B. Sample Use Case – Coordinator Agent rewriting the theory and adding the respective superiority
relationships, according to the scope of each rule set. Thus, for
The preference rule sets above are of no utility by themselves, the specific rule set presented in this subsection, the following
since they have to be approved/rejected by the CA. The latter superiority relationships are appended to the rule base:
is already equipped with its own policy, containing
maintenance and emergency rules: m1 > pA1-11
m1 > pA2-12
# maintenance rules e1 > m1
m0: consumption(X, Y), Y > 2000 → savingMode(X)
After every update in its knowledge base, the CA executes
m1: savingMode(X) ⇒ switchOff(X, light) the rule set through SPINdle [12], a state-of-the-art defeasible
m2: isOn(X, light) x ¬switchOn(Χ, light) #defeater logic reasoning engine. The following two brief examples give
m3: isOff(Χ, light) x¬switchOff(Χ, light) #defeater a deeper insight into the reasoning process.
Example 1
# emergency rules
e0: smoke(X) → alert(X) Suppose that the following facts are inserted into the above
e1: alert(X) ⇒ switchOn(X, light) rule base:
As can be observed, maintenance rules tend to enforce energy f1: time(2300)
savings, often compromising the comfort requested by the f2: consumption(A1,2300)
user-defined preference rules. The specific rule set determines Fact f1 triggers rules pA1-01 and pA2-02 which, in turn, trigger
the conditions for having each room switch to ‘saving mode’, rules pA1-11 and pA2-12 that conclude switchOn(A1, light) and
during which the system will attempt to minimize switchOn(A2, light). On the other hand, fact f2 triggers rule m0
consumption. Also, the defeaters make sure that the light’s that, in turn, triggers rule m1, which concludes switchOff(A1,
actuator will not attempt to turn on an already operating light light). Rules m1 and pA1-11 are conflicting and this conflict is
source and, conversely, will not attempt to turn off a light resolved via the superiority relationship m1 > pA1-11, which
source that has already been switched off. Additionally, the eventually concludes that the light in room A1 should be turned
emergency rules handle situations that imply unsafe conditions off. The light in room A2, however, remains turned on.
inside a room, like e.g. smoke. In this case, the system Example 2
activates the room’s alert, which then handles the operation of
the appliances in the room accordingly (e.g. lights are turned Suppose now that fact f3: smoke(A1) is also added into the
on to help users find their way out of the room). knowledge base. As a consequence, rule e0 will be triggered,
which, in turn, triggers rule e1 that concludes switchOn(A1,
Whenever a room policy is updated, the respective RA light). A new conflict arises, this time between e1 and m1, but,
immediately submits the revised rule set to the CA. In the since a respective superiority relationship has already been
running example, the following rules will be appended to the established (e1 > m1), the conflict is resolved, resulting in the
CA’s policy (notice that RA rules undergo certain syntactic lights being eventually turned on.
modifications when merged into the CA rule base):
pA1-01: time(X), X > 1600 → afternoon IV. STATE OF THE ART
pA1-02: time(X), X < 1600 → ¬afternoon As far as rule-based smart environments are concerned, one
pA1-11: afternoon ⇒ switchOn(A1, light) approach introduced in [13] is a meta-language defined over
pA1-12: ¬ afternoon ⇒ switchOff(A1, light)
pA2-01: time(X), X < 1700 → dayTime

129
JESS10, to syntactically enhance the rule authoring process in instantly be modified accordingly. This functionality fits our
ambient applications. However, this additional syntactic layer, requirements and will potentially improve the flexibility of our
called Event-Control-Action (ECA) model, is far less flexible proposed framework.
and extensible than defeasible logic used in the proposed
framework. Other similar approaches include SESAME-S [7], ACKNOWLEDGMENT
an all-in-a-box smart home prototype that uses ontologies and
JESS reasoning to enforce rules. The main drawback is the lack The authors would like to thank Konstantinos Gottis, John
of conflict resolution and scalability, while our work targets Argyriou, George Pilikidis, Thodoris Tsompanidis and Andrea
those aspects with defeasible logic. Regarding the proposed Dimitri for their work in various parts of system integration.
middleware approach, the SAWSDL bottom-up, lightweight
standard is used, while most existing paradigms use more REFERENCES
complex and, thus, less interoperable top-down standards, as in [1] E. I. A. (US), Annual Energy Review 2011. Government Printing Office,
[10] [14], or custom ones, as in [15] [16]. 2012.
[2] U. S. DoE, “Buildings energy data book,” Energy Effic. Renew. Energy
Multi-agent systems have also been used for managing Dep., 2011.
energy consumption of buildings. The approaches presented in [3] T. Berners-Lee, J. Hendler, and O. Lassila, “The semantic web,” Sci.
[17] and [18] represent various entities in buildings as agents Am., vol. 284, no. 5, pp. 28–37, 2001.
and target the reduction of the building’s consumption, while [4] D. J. Cook, J. C. Augusto, and V. R. Jakkula, “Ambient intelligence:
Technologies, applications, and opportunities,” Pervasive Mob.
comfort factors of the occupants are also taken into Comput., vol. 5, no. 4, pp. 277–298, 2009.
consideration. More specifically, [17] employs multi-objective [5] V. Issarny, M. Caporuscio, and N. Georgantas, “A perspective on the
Markov Decision Problems, and [18] uses Particle Swarm future of middleware-based software engineering,” in 2007 Future of
Optimization in order to optimally schedule energy consuming Software Engineering, 2007, pp. 244–258.
activities. On the contrary, our work provides agent [6] T. G. Stavropoulos, K. Gottis, D. Vrakas, and I. Vlahavas, “aWESoME:
coordination and defeasible logics for conflict resolution, while A web service middleware for ambient intelligence,” Expert Syst. Appl.,
vol. 40, no. 11, pp. 4380–4392, 2013.
scheduling is encoded in policies authored by experts. [7] L. Daniele, P. D. Costa, and L. F. Pires, “Towards a rule-based
approach for context-aware applications,” in Dependable and Adaptable
V. CONCLUSIONS AND FUTURE WORK Networks and Services, Springer, 2007, pp. 33–43.
[8] T. G. Stavropoulos, D. Vrakas, D. Vlachava, and N. Bassiliades,
The proposed multi-agent coordination framework is able “BOnSAI: a smart building ontology for ambient intelligence,” in
to optimize a building’s appliances to achieve long-term energy Proceedings of the 2nd International Conference on Web Intelligence,
savings while providing comfort. It does so, based on Mining and Semantics, 2012, p. 30.
[9] D. Nute, “Defeasible logic,” Handb. Log. Artif. Intell. Log. Program.,
preference, maintenance and emergency policies authored by vol. 3, pp. 353–395, 1994.
experts using human-intuitive defeasible logics and semantics. [10] E. Kontopoulos, T. Zetta, and N. Bassiliades, “Semantically-enhanced
The framework is based on an existing Semantic Web Service authoring of defeasible logic rule bases in the semantic web,” in
and wireless sensor and actuator network deployment at a Proceedings of the 2nd International Conference on Web Intelligence,
university building. Admittedly, the presented implementation Mining and Semantics, 2012, p. 56.
[11] D. Billington, “Conflicting literals and defeasible logic,” in Proc. 2nd
is at a preliminary stage, but additional extensions and Australian Workshop on Commonsense Reasoning, 1997, pp. 1–14.
experimentation are well underway. [12] H.-P. Lam and G. Governatori, “The making of SPINdle,” in Rule
Interchange and Applications, Springer, 2009, pp. 315–322.
A crucial future task to accomplish is to implement the [13] M. Eisenhauer, P. Rosengren, and P. Antolin, “A development platform
proposed framework on the target deployment, using existing for integrating wireless devices and sensors into ambient intelligence
components. Namely, communication and information systems,” in Sensor, Mesh and Ad Hoc Communications and Networks
exchange between existing agents needs to be established. Workshops, 2009. SECON Workshops’ 09. 6th Annual IEEE
Consequently, energy-saving results from such an application Communications Society Conference on, 2009, pp. 1–3.
[14] G. Thomson, S. Bianco, S. B. Mokhtar, N. Georgantas, and V. Issarny,
need to be measured, using the existing monitoring
“Amigo aware services,” in Constructing Ambient Intelligence,
infrastructure, especially, during late spring and summer Springer, 2008, pp. 385–390.
months when a power-demanding cooling system is activated. [15] A. Fensel, S. Tomic, V. Kumar, M. Stefanovic, S. V. Aleshin, and D. O.
Novikov, “Sesame-s: Semantic smart home system for energy
Another interesting future direction would be to build our efficiency,” Inform.-Spektrum, vol. 36, no. 1, pp. 46–57, 2013.
proposed model based on an advanced multi-agent framework [16] S. M. Iacob, J. P. A. Almeida, and M. E. Iacob, “Optimized dynamic
like EMERALD [19]. The latter allows defining interoperating, semantic composition of services,” in Proceedings of the 2008 ACM
knowledge-based intelligent agents, each of which is equipped symposium on Applied computing, 2008, pp. 2286–2292.
with a reasoning engine forming a shared reasoning [17] L. Klein, J. Kwak, G. Kavulya, F. Jazizadeh, B. Becerik-Gerber, P.
Varakantham, and M. Tambe, “Coordinating occupant behavior for
infrastructure. Although EMERALD currently features a few building energy and comfort management using multi-agent systems,”
but widely diverse reasoning services, those services can easily Autom. Constr., vol. 22, pp. 525–536, 2012.
be extended. A further advantage of the framework is that [18] Z. Wang, R. Yang, and L. Wang, “Multi-agent control system with
agents are knowledge-customizable, meaning that they are not intelligent optimization for smart and energy-efficient buildings,” in
confined in having their logics and strategies/policies hard- IECON 2010-36th Annual Conference on IEEE Industrial Electronics
Society, 2010, pp. 1144–1149.
wired. Instead, they can be customized by altering the rule [19] K. Kravari, E. Kontopoulos, and N. Bassiliades, “EMERALD: a multi-
base, while the agent’s knowledge and/or behavior will agent system for knowledge-based reasoning interoperability in the
semantic web,” in Artificial Intelligence: Theories, Models and
10 Applications, Springer, 2010, pp. 173–182.
http://herzberg.ca.sandia.gov/

130