Received April 23, 2018, accepted June 25, 2018, date of publication xxxx 00, 2018, date of current version xxxx 00, 0000. Digital Object Identifier 10.1109/ACCESS.2017.DOI A Game Theoretical Randomized Method for Large-Scale Systems Partitioning FRANCISCO JAVIER MUROS 1 , JOSÉ MARÍA MAESTRE 1 , (Senior Member, IEEE), CARLOS OCAMPO-MARTINEZ 2 , (Senior Member, IEEE), ENCARNACIÓN ALGABA 3 , and EDUARDO F. CAMACHO 1 , (Fellow, IEEE) 1 Department of Systems and Automation Engineering, University of Seville, Spain (e-mail: {franmuros,pepemaestre,efcamacho}@us.es) 2 Automatic Control Department, Universitat Politècnica de Catalunya, Institut de Robòtica i Informàtica Industrial (CSIC-UPC), Barcelona, Spain (e-mail: [email protected]) 3 Department of Applied Mathematics II, University of Seville, Spain (e-mail: [email protected]) Corresponding author: Francisco Javier Muros (e-mail: [email protected]) This work has been partially supported by the European Union Project FP7-ICT-ICT-2013.3.4-611281 (DYMASOS), and the MINECO-Spain Projects DPI2016-78338-R (CONFIGURA), DPI2017-86918-R, ECO2015-68856-P and DPI2016-76493-C3-3-R (DEOCS). ABSTRACT In this paper, a game theory-based partitioning algorithm for large-scale systems (LSS) is proposed. More specifically, a game over nodes is introduced in a model predictive control framework. The Shapley value of this game is used to rank the communication links of the control network based on their impact on the overall system performance. A randomized method to estimate the Shapley value of each node and also an efficient redistribution of the resulting value to the links involved are considered to relieve the combinatorial explosion issues related to LSS. Once the partitioning solution is obtained, a sensitivity analysis is proposed to give a measure of its performance. Likewise, a greedy fine tuning procedure is considered to increase the optimality of the partitioning results. The full Barcelona drinking water network (DWN) is analyzed as a real LSS case study showing the effectiveness of the proposed approach in comparison with other partitioning schemes available in the literature. INDEX TERMS Coalitional control, cooperative game theory, system partitioning, randomized methods, Shapley value, large-scale systems (LSS), drinking water networks (DWN). I. INTRODUCTION M ODEL predictive control (MPC) has evolved consid- erably over the last decades. It designates an ample range of control methods that make explicit use of a model of the process to obtain the control signal by minimizing an objective function [1]. Its flexibility – for any type of model can be used – and the ease of dealing with constraints and dead times are well-known advantages of this methodology. In particular, this paper focuses on distributed MPC (DMPC) schemes, where the overall control problem is divided into smaller pieces assigned to local controllers or agents, which are able to communicate among them. Typical features of these approaches, such as scalability, modularity, and the capacity of controlling large-scale systems (LSS) [2], will be very welcome in this work. Each agent involved in a distributed scheme has partial system information and communicates with neighbors for the sake of coordination [3]. Typically, the partition of the overall system is assumed to be given before the DMPC strategy is applied, and it is calculated during the system modeling, based on physical insight, experience or intuition, and other methods available in the literature, where the starting point is commonly associated with the seminal work of Siljak [4]. Recently, many partitioning schemes have appeared, based on graph theory [5]–[8], states/inputs estimation [9], social network algorithms [10], genetic algorithms [11], or PageR- ank [12]. In any case, no single partitioning strategy is the best fit for all situations. This way, specific partitioning tech- niques have been applied to real LSS case studies, e.g., in wa- ter systems [10], [13]–[15], power networks [16], [17], bio- logical systems [18], integrated circuits [19], and urban traffic networks [20]. In fact, LSS might involve a big communica- tion network implying the handling of large amount of data, which could yield in high computational costs. Therefore, performing the partition of the problem into smaller pieces is a natural solution for managing these networks. VOLUME 4, 2016 1
19
Embed
A Game Theoretical Randomized Method for Large-Scale ...
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Received April 23, 2018, accepted June 25, 2018, date of publication xxxx 00, 2018, date of current version xxxx 00, 0000.
Digital Object Identifier 10.1109/ACCESS.2017.DOI
A Game Theoretical Randomized Methodfor Large-Scale Systems PartitioningFRANCISCO JAVIER MUROS 1, JOSÉ MARÍA MAESTRE 1, (Senior Member, IEEE),CARLOS OCAMPO-MARTINEZ 2, (Senior Member, IEEE), ENCARNACIÓN ALGABA 3,and EDUARDO F. CAMACHO 1, (Fellow, IEEE)1Department of Systems and Automation Engineering, University of Seville, Spain (e-mail: {franmuros,pepemaestre,efcamacho}@us.es)2Automatic Control Department, Universitat Politècnica de Catalunya, Institut de Robòtica i Informàtica Industrial (CSIC-UPC), Barcelona, Spain(e-mail: [email protected])3Department of Applied Mathematics II, University of Seville, Spain (e-mail: [email protected])
Corresponding author: Francisco Javier Muros (e-mail: [email protected])
This work has been partially supported by the European Union Project FP7-ICT-ICT-2013.3.4-611281 (DYMASOS),and the MINECO-Spain Projects DPI2016-78338-R (CONFIGURA), DPI2017-86918-R, ECO2015-68856-P andDPI2016-76493-C3-3-R (DEOCS).
ABSTRACT In this paper, a game theory-based partitioning algorithm for large-scale systems (LSS) isproposed. More specifically, a game over nodes is introduced in a model predictive control framework.The Shapley value of this game is used to rank the communication links of the control network based ontheir impact on the overall system performance. A randomized method to estimate the Shapley value ofeach node and also an efficient redistribution of the resulting value to the links involved are consideredto relieve the combinatorial explosion issues related to LSS. Once the partitioning solution is obtained,a sensitivity analysis is proposed to give a measure of its performance. Likewise, a greedy fine tuningprocedure is considered to increase the optimality of the partitioning results. The full Barcelona drinkingwater network (DWN) is analyzed as a real LSS case study showing the effectiveness of the proposedapproach in comparison with other partitioning schemes available in the literature.
INDEX TERMS Coalitional control, cooperative game theory, system partitioning, randomized methods,Shapley value, large-scale systems (LSS), drinking water networks (DWN).
I. INTRODUCTION
MODEL predictive control (MPC) has evolved consid-erably over the last decades. It designates an ample
range of control methods that make explicit use of a modelof the process to obtain the control signal by minimizing anobjective function [1]. Its flexibility – for any type of modelcan be used – and the ease of dealing with constraints anddead times are well-known advantages of this methodology.In particular, this paper focuses on distributed MPC (DMPC)schemes, where the overall control problem is divided intosmaller pieces assigned to local controllers or agents, whichare able to communicate among them. Typical features ofthese approaches, such as scalability, modularity, and thecapacity of controlling large-scale systems (LSS) [2], will bevery welcome in this work.
Each agent involved in a distributed scheme has partialsystem information and communicates with neighbors for thesake of coordination [3]. Typically, the partition of the overall
system is assumed to be given before the DMPC strategyis applied, and it is calculated during the system modeling,based on physical insight, experience or intuition, and othermethods available in the literature, where the starting pointis commonly associated with the seminal work of Siljak [4].Recently, many partitioning schemes have appeared, basedon graph theory [5]–[8], states/inputs estimation [9], socialnetwork algorithms [10], genetic algorithms [11], or PageR-ank [12]. In any case, no single partitioning strategy is thebest fit for all situations. This way, specific partitioning tech-niques have been applied to real LSS case studies, e.g., in wa-ter systems [10], [13]–[15], power networks [16], [17], bio-logical systems [18], integrated circuits [19], and urban trafficnetworks [20]. In fact, LSS might involve a big communica-tion network implying the handling of large amount of data,which could yield in high computational costs. Therefore,performing the partition of the problem into smaller piecesis a natural solution for managing these networks.
F. J. Muros et al.: A Game Theoretical Randomized Method for Large-Scale Systems Partitioning
In this work, the coalitional control framework, which canbe summarized as a set of dynamic partitioning methods fornetworked control systems that seek for a trade-off betweencommunication burden and control performance [21]–[26],is used to determine static neighborhoods that define thepartitioning of an LSS. That is, despite its dynamic scope,the coalitional framework can be used to obtain offline systeminformation, i.e., independent from the implementation of thecontrol scheme. In this coalitional context, the well-knownShapley value, a cooperative game solution concept pre-sented originally in [27], is used here to provide informationregarding the relevance of the agents and the links involvedin a communication network, following the ideas describedin [21]–[23]. The specific contributions of this article withrespect these earlier works are the following:
a) A game defined in the set of agents is proposed here,where the closed-loop stage cost of the system is ac-cumulated along the simulation scenario. This is a dif-ference with respect to previous works, were open-loopcosts where used at each time step. This game is associ-ated with the cost function of a coalitional MPC scheme,in the line of [23]. Note that the viewpoint is changedwith respect to [21], [22], where a game with the playersbeing the links was related to the cost function of acoalitional approach based on linear feedback gains.
b) The Shapley value, which gives an averaged contribu-tion of each player into the game, is calculated hereby using randomized methods [28], [29] satisfying aminimum bounded error requirement. The relevance ofthe links is obtained by an index, introduced in [23],which redistributes the Shapley value from the agentsto the links.
c) In [21], [23], a very preliminar partitioning that usesfew thresholds to classify the links as a function of thecontrol performance is proposed. Here, the partitioningalgorithm is drastically improved by introducing newparameters that balance the size of the communicationcomponents. Additionally, a new set that refers to ex-pensive links among components is also introduced hereto provide a proper definition of the network configura-tion after applying the partitioning method.
d) A sensitivity analysis that gives a measure of the par-titioning control performance is included in this pa-per. Through this analysis, the proposed partitioningsolution is compared with other schemes available in theliterature [10], [13], and also with the centralized anddecentralized configurations. Moreover, this analysis isrecursively implemented in a greedy fashion [30], [31]to make a fine tuning of the partitioning approach.
Notice that the results reported in previous works wereonly suitable for academic small networks. Nevertheless,through the combination of a) and b) the computational andcombinatorial explosion issues related to LSS are mitigatedand it is possible to implement the new partitioning algorithmintroduced in c) and d) in such networks. Moreover, the full
Barcelona drinking water network (DWN), modeled in [13],[14], is chosen as a real LSS case study to assess the effective-ness of the proposed partitioning approach, which representsan additional contribution of this paper.
The remainder of this paper is organized as follows. InSection II, the problem setting is stated in a coalitionalcontrol framework. Next, in Section III, a game over nodesis considered, being the Shapley value utilized as a means ofdistributing the cost of the game among the agents. Likewise,a randomized method to estimate the Shapley value and aredistribution of this value to the set of links are proposedto deal with combinatorial explosion issues. In Section IV,a Shapley value-based partitioning algorithm is introduced.The partitioning performance is evaluated by means of asensitivity analysis, whose information is also recursivelyused for a fine tuning of the proposed approach. In Section V,the Barcelona DWN is presented as the case study, withthe corresponding partitioning results and comparisons withother approaches being presented in Section VI. Finally,conclusions are drawn in Section VII.
II. FOUNDATIONS OF COALITIONAL CONTROLConsider the class of distributed linear systems composed ofN = {1, 2, . . . , |N |} interconnected subsystems or agents.The dynamics of agent i ∈ N can be mathematicallydescribed, with k ∈ Z+, as
xi(k + 1) = Aiixi(k) + Biiui(k) + wi(k),
wi(k) =∑j 6=i
[Aijxj(k) + Bijuj(k)] + Bpidi(k),
(1)where xi(k) ∈ Rnxi is the state vector of agent i, ui(k) ∈Rnui its corresponding input vector, and wi(k) ∈ Rnxi
the related disturbances, which can be either external to thewhole system, denoted by di(k) ∈ Rndi , or be caused bythe neighbors as well. Likewise, Aii ∈ Rnxi
×nxi ,Bii ∈Rnxi
×nui , Aij ∈ Rnxi×nxj ,Bij ∈ Rnxi
×nuj and Bpi ∈Rnxi
×ndi are system matrices of suitable dimensions. Bothstates and inputs are constrained into independent sets de-fined by a collection of linear inequalities, i.e.,
A. NETWORKED COALITIONAL STRUCTUREIn standard coalitional control, the agents are merged at eachtime instant into several disjoint neighborhoods or commu-nication components C1, C2, . . . , Cnc , verifying
⋃ncs=1 Cs =
N . Conversely, the goal in this paper is to use coali-tional control to find a time-independent set denoted asNC = {C1, C2, . . . , Cnc}, i.e., nc = |NC |, assuming thatagents inN are initially connected by a network described byan undirected graph (N , E), where E = EN = N ×N is theset of links corresponding to all feasible communication con-nections among the agents. Hence, the number of elements inboth sets is, in the worst case, connected by [21], [23]
|E| = |N |(|N | − 1)
2, (3)
2 VOLUME 4, 2016
F. J. Muros et al.: A Game Theoretical Randomized Method for Large-Scale Systems Partitioning
which corresponds to the number of links of a completenetwork. This case, which is the most demanding in termsof problem size, will be assumed throughout the paper. Notethat, under this premise, any group or coalition C ⊆ N ofagents is internally connected.
Each link l ∈ E can be classified according to its relevancefrom a control viewpoint. This way, it can be more profitablefor the overall system performance to fix/disconnect somelinks permanently. In this work, the partitioning objectivewill focus on finding out those links. When a specific coali-tion C is formed, a model analogous to (1) is calculated at acoalition level, i.e.,
xC(k + 1) = ACCxC(k) + BCCuC(k) + wC(k),
wC(k) =∑j /∈C
[ACjxj(k) + BCjuj(k)] + BpCdC(k),
(4)with xC(k) = [xi(k)]i∈C ∈ RnxC , uC(k) = [ui(k)]i∈C ∈RnuC and wC(k) = [wi(k)]i∈C ∈ RnwC , dC(k) =[di(k)]i∈C ∈ RndC being respectively the coalitional states,inputs, overall disturbances and external disturbances thataggregate the corresponding vectors, and ACC ∈ RnxC×nxC ,BCC ∈ RnxC×nuC , ACj ∈ RnxC×nxj , BCj ∈ RnxC×nuj
and BpC ∈ RnxC×ndC are obtained by aggregating the cor-responding individual matrices. The coalitional constraintsbecome
xC(k) ∈ XC ⊆ RnxC , XC =∏i∈C Xi,
uC(k) ∈ UC ⊆ RnuC , UC =∏i∈C Ui.
(5)
Finally, from an overall centralized viewpoint, the system isdescribed by
xN (k+ 1) = ANxN (k) +BNuN (k) +BpNdN (k), (6)
with xN (k) ∈ RnxN , uN (k) ∈ RnuN , dN (k) ∈ RndN
being, respectively, the overall state, input and disturbancevectors. Note that from a centralized viewpoint, wN (k) isonly composed of external disturbances BpNdN (k). Thecentralized constraints have the form
xN (k) ∈ XN ⊆ RnxN , XN =∏i∈N Xi,
uN (k) ∈ UN ⊆ RnuN , UN =∏i∈N Ui.
(7)
B. CONTROL OBJECTIVEUnder an MPC framework, the goal of each coalitionalcontroller C ⊆ N is to drive a sequence of future states overa prediction horizon Np, that is, XC(k + 1 : k + Np) ={xC(k + 1), . . . ,xC(k + Np)}, by using the most appropri-ate control sequence. To this end, the controller solves thefollowing open-loop finite-horizon optimization problem ateach time instant k:
U∗C(k : k +Np − 1) =
arg minUC(k:k+Np−1)
Np−1∑r=0
`C(xC(k + r + 1),uC(k + r)
),
(8)
subject to (4), (5), the aggregate forecast of the expected localdisturbances WC(k : k + Np − 1) = {wC(k), . . . , wC(k +Np − 1)}, and a measured coalitional initial state xc(k).Likewise, `C
(xC(k),uC(k)
)is a certain convex coalitional
stage cost that is minimized along the prediction horizon. Asa result, the sequence of the coalitional optimal control inputsover Np, that is, U∗C(k : k+Np− 1) = {u∗C(k), . . . ,u∗C(k+Np − 1)} is obtained. Only the first control input u∗C(k)is actually applied, and the rest of elements are discarded.At the next time instant, (8) is solved again in a recedinghorizon fashion.
Notice that a coalition can be a singleton. In this case,each local controller i ∈ N solves an optimization problemanalogous to (8) by taking C = i. Likewise, to computea centralized MPC scheme implemented in a distributedfashion it is enough to calculate the optimal input sequenceby taking C = N and solving (8).
III. COOPERATIVE GAMES AND THE SHAPLEY VALUE
In [21], [22], the set of links E was interpreted as the set ofplayers of a cooperative game whose characteristic functionassigned a value to each of the different configurations oflinks or network topologies. As commented before, in theapproach proposed this paper it is assumed that the number oflinks is related to the agents by (3). Hence, to mitigate com-binatorial explosion issues associated to LSS, the perspectiveof the aforementioned works is changed here, as done in [23],to working directly with a cooperative game (N ,v) definedover the set of agents N . To this end, a cost function vthat assigns a cost to each coalition of players C ⊆ N isdefined by
v(C,xN ) =Tsim−1∑k=0
[`C(xC(k + 1),u∗C(k)
)+∑i/∈C
`i(xi(k + 1),u∗i (k)
)],
(9)with `i
(xi(k),ui(k)
)and `C
(xC(k),uC(k)
)being the stage
costs, which will be defined for the case study in Section V-B,and where Tsim is the number of simulation steps used toaccumulate the closed-loop stage cost of the system over thesimulation time. This function v is computed by applying ateach time step the first element of the control sequence ofcoalition C, i.e., u∗C(k), which is obtained by solving (8). Therest of the agents calculate their input sequences u∗i (k) bysolving (8), with C = i, independently.
Remark 1. Equation (9) is evaluated with input informationfrom all agents, independently whether they are either inor out the coalition C. Each coalition C solves its ownoptimization problem, which is decoupled from the rest of thenetwork. Hence, 2|N | values are needed to fully characterizefunction v using the standard approach, which is not the casein this work, as it will be shown in the next subsection byconsidering randomized methods.
VOLUME 4, 2016 3
F. J. Muros et al.: A Game Theoretical Randomized Method for Large-Scale Systems Partitioning
Once the game is defined, the Shapley value [27] is con-sidered here to get the corresponding cost that each playerexpects from the game. This value assigns to game (N ,v)vector φ(N ,v), defined ∀i ∈ N as
φi(N ,v) =∑C⊆N ,i/∈C
|C|!(|N | − |C| − 1)!
|N |![v(C ∪ {i},xN )− v(C,xN )],
(10)i.e., the marginal contribution of each agent i ∈ N isaveraged for all possible coalition permutations it can bepart of.
Remark 2. Equation (10) was originally defined in thecontext of transferable utility (TU) games. Given that agentsout of C work independently, an univocal v(C) is obtainedfor each C, and (9) could be interpreted as a TU-game bysimply assuming a redefined game v′(C,xN ) = v(C,xN ) −v(∅,xN ). This way, the Shapley value of this redefined game,which trivially coincides with the Shapley value of the origi-nal game (see [22]), will be considered here.
Given that the partitioning procedure proposed in thispaper will be performed by fixing/disconnecting some com-munication links among the different agents, a measure of therelevance of the links is required. Given a link l = {i, j} ∈ E ,it is possible to redistribute the Shapley value of the agentsthat are the end-points of this link, i.e., i and j, by means ofthe following expression [23]:
ξl(N ,v) =1
|Ei|φi(N ,v) +
1
|Ej |φj(N ,v), (11)
with Ei and Ej being, respectively, the set of links connectedto agents i and j. Notice that values ξl(N ,v) satisfy effi-ciency as the original Shapley value, i.e.,∑
l∈E
ξl(N ,v) =∑i∈N
φi(N ,v) = v(N ,xN ). (12)
Note that (11) provides a way to arrange and compare thelinks according to their relevance from a control-performanceperspective, which can be interpreted as a LinkRank, in theline of [32]. This way, the lower this value is, the more usefulthe link becomes. This is consistent with the Shapley value,which associates useful players to lower values.
Remark 3. Equation (11) provides information from all linksl ∈ E by means of the Shapley value of agent-based game (9),which is obtained by evaluating the corresponding 2|N |
coalitions. This fact mitigates the combinatorial explosionof the method proposed in [21], which depends on a link-game that requires to evaluate 2|E| coalitions per game, with|E| = f(|N |2) as shown in (3).
A. ESTIMATION OF THE SHAPLEY VALUEIn networks with a large number of agents it is not computa-tionally feasible to compute (9) for every coalition. This issuecan be solved by using randomized methods such as [28],
[29]. In particular, the method in [28] is used here to providean estimation of the Shapley value of each agent calculatedin polynomial time provided that (9) can also be calculatedin polynomial time, which will be shown in Section V-B. Tothis end, the following alternative definition of the Shapleyvalue, expressed in terms of all possible orderings of playersinN coming into coalition, i.e., |N |!, is used in the samplingmethod given in [28]:
φi(N ,v) =1
|N |!∑
π∈Π(N )
mπi (N ,v), ∀i ∈ N , (13)
where the orderings are assumed to be equiprobable, withΠ(N ) being the collection of all permutations π, and where
mπi (N ,v) = v ({j ∈ N : π(j) ≤ π(i)})
− v({j ∈ N : π(j) < π(i)}),(14)
is the marginal contribution of player i to the players that areranked before it in permutation π.
The basics of the method in [28] consist in choosinga given number of random orderings from set Π(N ) forestimating the Shapley value of each player. To this end, aset Q that contains a sample of q different permutations π,which are taken with replacement and with equal probabilityfrom set Π(N ), is considered. This way, an estimation ofthe Shapley value is given by the average of the marginalcontributions over set Q, i.e.,
φi(N ,v) =1
q
∑π∈Q
mπi (N ,v), ∀i ∈ N . (15)
Equation (15) provides an approximation of the Shapleyvalue with desirable properties. In particular, as the Shapleyvalue, the estimator satisfies efficiency. Moreover, followingthe central limit theorem, it holds [28] that the estimatoris a normal distribution with the following mean value andstandard deviation:
φi(N ,v) ∼ N
(φi,
σ2φi
q
), (16)
with
σ2φi =
1
|N |!∑
π∈Π(N )
(mπi (N ,v)− φi(N ,v)
)2, ∀i ∈ N .
(17)Consequently, if the number of permutations q is chosensatisfying the following condition, ∀i ∈ N [28]:
q ≥Z2λ/2σ
2φi
ε2, (18)
the estimation error is is guaranteed to be bounded by [28]
P(|φi(N ,v)− φi(N ,v)| ≤ ε
)≥ 1− λ, (19)
with ε being the approximation error, Z ∼ N(0, 1), andwhere Z2
λ/2 is the value such that P (Z ≥ Z2λ/2) = λ/2,
with 0 ≤ λ ≤ 1. Finally, note that by taking σφi = αε,i.e., proportionally to the desired error, condition (18) isreduced to
q ≥ αZ2λ/2. (20)
4 VOLUME 4, 2016
F. J. Muros et al.: A Game Theoretical Randomized Method for Large-Scale Systems Partitioning
IV. SHAPLEY-VALUE-BASED PARTITIONINGALGORITHMThe main objective of the algorithm presented in this paper isto find which agents should cooperate to improve the overallsystem performance. Notice that, some preliminar steps areneeded before it can be launched. In the first place, there is apre-partitioning stage in which a set of atomic agents needsto be defined by assigning states, actions and constraints tominimum size entities that could work in a fully decentralizedfashion, i.e., without the need for communication with otherparts of the system. In this work, as it will be shown inSection V-A, these agents stem from the constraints imposedon the system by the node equations.
Once the agents are defined, a communication link be-tween each pair of agents is considered, with the full numberof links given by (3). Then, it is needed to classify theselinks according to their relevance in terms of their impacton the overall system performance. To this end, a measureindex related to the Shapley value is associated with eachlink as follows:
Measure Indices ProcedureLet (N , E) be a network that describes a set N of agents
connected by links l ∈ E . Consider also a measured initialstate xN (k) and a forecast of the expected disturbancesWN (k : k +Np − 1). Then,
a) Calculate a size q that guarantees, following (18), thatthe estimation error is under desired limits.
b) Compute φi(N ,v),∀i ∈ N , by using (15). For eachcoalition C ∈ NC , the optimal input sequence over Npis obtained by solving (8) for C and also for the agentsout of C. Only the first control input is applied, and therest of elements are discarded. At the next time stepthese optimization problems are solved in a recedinghorizon fashion. This process is performed during Tsimtime instants and the cumulated cost of this closed-loopsimulation is used to build v(C,xN ) by means of (9).
c) Redistribute the obtained Shapley value among the linksby indices ξl(N ,v), calculated by using (11).
Therefore, index ξl(N ,v) measures the impact of link l inthe control network. Based on these indices, it is possible torank the links, which in turns allows to obtain the followingsubsets:
• Set Ec ⊆ E : it includes the links that are inexpensiveenough in control terms to always be fixed. This way,agents connected by links belonging to Ec cooperatetogether and will be merged in a single agent, whichcorresponds to any of the communication componentsin NC = {C1, C2, . . . , Cnc}.
• Set Ee ⊆ E : it comprises the links that are too costlyfor the system in terms of control performance, andtherefore will be permanently disconnected.
Basically, the partitioning algorithm in this section is intro-duced with the aim of determining sets Ec and Ee. Both setsare initially assumed to be empty, and links will be graduallyincluded in those sets if they satisfy certain conditions. Thisway, links in Ec(s) define the communication components inNC(s) in a given iteration s ∈ N+. Some concepts that willbe needed to perform the partitioning are introduced next:• K: symbolizes the mean span among indices ξl(N ,v)
and it is defined by
K =ξlmax(N ,v)− ξlmin(N ,v)
|E|, (21)
where lmax and lmin denote, respectively, the links withmaximum and minimum measure indices ξl(N ,v).
• Ci(s): denotes the component in NC(s) where agent ibelongs to.
• Ei(s): denotes the set of remaining links in E\Ec(s)connected to agent i.
Notice that the sizes of sets Ci(s) and Ei(s) are inverselyrelated. Both sets will be of interest in the Shapley-value-based partitioning algorithm, which is presented below.
F. J. Muros et al.: A Game Theoretical Randomized Method for Large-Scale Systems Partitioning
Partitioning Algorithm 1
Let ξl(N ,v) be the indices related to each link l ∈ E . LetLc, Le ∈ R be given thresholds, verifying Lc ≤ Le. Finally,let s be a counter variable.
I) Computation of set Ec
Let γ, ρ ∈ R+\{0}, ν ≥ 1 be scalar parameters. Let alsoK be the mean span given by (21). Starting with s = 0, andassuming ξ(0)
l (N ,v) = ξl(N ,v) and Ec(0) = ∅, do
1) Obtain the link with minimum measure index from theremaining links that are not yet included in set Ec(s),i.e.,
l∗c = arg minlξ
(s)l (N ,v), l ∈ E\Ec(s). (22)
2) Validate link l∗c = {i∗, j∗} as a suitable candidate tobe added to Ec(s), and update this set and the linksneighbors of l∗c consequently, doing, ∀l ∈ E\Ec(s + 1),the procedure defined by (23), with Γ being calculatedas
Γ = γK (|Ci∗(s)|+ |Cj∗(s)|)ν . (24)
3) Make s = s + 1 and go to Step 1, while ξ(s)l∗c
(N ,v) <Lc. Otherwise, the procedure ends and set Ec is fullydetermined, i.e., Ec = Ec(s).
II) Computation of set Ee
Starting again with s = 0, and assuming ξ(0)l (N ,v) =
ξl(N ,v), Ee(0) = ∅, do4) Obtain the link with maximum measure index from the
remaining links, i.e.,
l∗e = arg maxlξ
(s)l (N ,v), l ∈ E\(Ec ∪ Ee(s)). (25)
5) Validate link l∗e = {i∗, j∗} as a suitable candidate forset Ee(s), and update this set consequently, doing
Ee(s+ 1) = Ee(s) ∪ l∗e , if ξ(s)l∗e
(N ,v) > Le,
Ee(s+ 1) = Ee(s), otherwise.(26)
6) Make s = s + 1 and go to Step 4, while ξ(s)l∗e
(N ,v) >Le. Otherwise, the procedure ends and set Ee is fullydetermined, i.e., Ee = Ee(s).
Notice that the inclusion of a new link l∗c in Ec(s) dependson the size of the components at s that this link will connect.More specifically, Γ penalizes a new link candidate to Ec(s)before deciding whether it should be included in that set, incase that this link would connect two different components inNC(s), and proportionally to their cardinality. Additionally,in case that l∗c is accepted, the proposed algorithm penalizesits neighboring links that remain in set E\Ec(s + 1), by aterm that is larger as less neighbors of l∗c remain in thatset, which in turns implies that l∗c is more congested inEc(s + 1). These mechanisms induce size constraints on the
FIGURE 1. Two iteration steps in the process of obtaining set Ec
components, which avoid inefficient partitionings that couldlead to an almost centralized scheme. This way, the proposedparameters γ, ν and ρ could be adjusted to obtain someproperties of interest, e.g., to impose a maximum cardinalityfor any communication component.
Example 1. Take the network in Fig. 1, with six agentsconnected by 15 links. In a given iteration s (see Fig. 1a),it is obtained
Ec(s) ={{1, 2}
}, NC(s) =
{{1, 2}, {3}, {4}, {5}, {6}
},
(27)
Then, consider that link l∗c = {i∗, j∗} = {1, 3} is proposedto be fixed. The sets related to this link in iteration s aredescribed by
(N ,v) + Γ < Lc. Thisway, given that C1(s) 6= C3(s), the first statement in (23)is fulfilled, hence link {1, 3} is included in set Ec(s+ 1) (seeFig. 1b), obtaining
Ec(s+ 1) ={{1, 2}, {1, 3}
},
NC(s+ 1) ={{1, 2, 3}, {4}, {5}, {6}
},
C1(s+ 1) = C3(s+ 1) = {1, 2, 3},E1(s+ 1) =
{{1, 4}, {1, 5}, {1, 6}
},
E3(s+ 1) ={{2, 3}, {3, 4}, {3, 5}, {3, 6}
}.
(29)
Therefore, ∀l ∈ E\Ec(s + 1), all measure indices areupdated if l belongs either to E1(s + 1) or E3(s + 1), byincreasing in an inversely proportional way to the cardinalityof those sets
ξ(s+1)l (N ,v) = ξ
(s)l (N ,v) + ρK
3, l ∈ E1(s+ 1),
ξ(s+1)l (N ,v) = ξ
(s)l (N ,v) + ρK
4, l ∈ E3(s+ 1).
(30)
Finally, consider that l∗c = {i∗, j∗} = {2, 3} is alsoproposed to be fixed. In that case, given that C2(s + 1) =C3(s + 1) = {1, 2, 3}, the second statement in (25) issatisfied, and the link will be included in set Ec(s + 2).Nevertheless, no updates on the measure indices of linksl ∈ E\Ec(s+ 2) would be considered.
6 VOLUME 4, 2016
F. J. Muros et al.: A Game Theoretical Randomized Method for Large-Scale Systems Partitioning
Once set Ec is fully determined, a new single agent cor-responding to each component in NC is established, inde-pendently of whether the agents inside this component aredirectly or indirectly connected. Therefore, the system isreduced to |NC | new agents, which may communicate or not,depending on the needs of the control scheme. The possiblelinks among the agents are denoted by EC = NC×NC , whichverifies |EC | = |NC|(|NC|−1)
2 , according to (3). That is, linkslC ∈ EC are defined as
lC = {Ca, Cb}, ∀Ca, Cb ∈ NC . (31)
Finally, the information provided by Ee is used to determinelinks lC that should be permanently disconnected. Given thatlinks in Ee are referred to agents instead of components,it is necessary to obtain a new set, say EeC , referred tocomponents. In this work, it is considered that two compo-nents should not have a direct cooperation if all links thatinterconnect the agents inside both components belong tosubset Ee, i.e.,
If l = {i, j} ∈ Ee,∀i ∈ Ca,∀j ∈ Cb −→ lC = {Ca, Cb} ∈ EeC .(32)
Summing up, the configuration of the system will bedescribed after performing the partitioning by the followingnetwork:
(NC , EC\EeC ), (33)
where the links in EC\EeC may be dynamically enabled ordisabled at each time instant by means of a coalitional controlapproach [22], [24].
Remark 4. The proposed Shapley-value-based algorithmrepresents a heuristic methodology for the system partition-ing that avoids an exhaustive exploration of every coalitioninvolved in the control scheme. Note that this methodologyis independent from the model dynamics or the game choice,i.e., nonlinear systems or alternative definitions for the gamedifferent to (9) are possible. For example, it might be consid-ered theoretical aspects such as stability or robustness [2],[33], which are out of the scope here since this work onlyfocus on the partitioning methodology. Likewise, for the sakeof clarity, the class of linear systems described by (1), whichis widely studied in the literature, is assumed here.
Remark 5. The way the agents inside a component are actu-ally connected once the partitioning is performed is beyondof the scope of this work. This issue could be dealt with,e.g., by using spanning tree algorithms [34], [35], in orderto find the minimum set of links that is necessary to connectall agents belonging to a given component. In any case, notethat the partitioning approach reduces the communicationcosts of the original centralized scheme, given that the agentsinside a component after performing the partitioning are onlyrequired to communicate to their neighbors.
FIGURE 2. Establishing the components and their removed links by Ec and Ee
Example 2. Consider again the network presented in Exam-ple 1, with six agents and 15 links. Assume that after applyingthe partitioning algorithm, sets Ec and Ee are given by (seeFig. 2a):
Ec ={{1, 2}, {1, 3}, {2, 3}, {4, 5}
},
Ee ={{1, 6}, {3, 6}, {4, 6}, {5, 6}
}.
(34)
The links in Ec define the following components:
C1 = {1, 2, 3}, C2 = {4, 5}, C3 = {6}, (35)
which are also assumed to be connected by links {C1, C2},{C1, C3} and {C2, C3}. From these three links, only link{C2, C3} verifies the criterion given in (32), i.e., all linksthat connect agents in components C2 and C3 are included inset Ee. Hence, this link should be permanently disconnected.Therefore, the final configuration of the network is shown inFig. 2b and it is given by
(NC , EC\EeC ) =({C1, C2, C3
},{{C1, C2}, {C1, C3}
}).(36)
A. SENSITIVITY ANALYSIS OF THE CONTROLPERFORMANCE: A PARTITIONING ALGORITHM FINETUNING
The rationale behind the partitioning procedure is to pro-vide a reasonable trade-off between control and communi-cation costs. Therefore, the proposed Shapley-value-basedapproach, denoted from now on by SVBA, provides us with asuboptimal solutionNC for the partitioning of an LSS, wherethe optimal solution corresponds with the centralized case(only one component) when communication costs are notconsidered. Once NC is established, the closed-loop systemperformance can be related to the cumulated cost Jcum(NC),which is obtained by computing each component C ∈ NCby using (9). Then, in order to give an insight of the SVBAfitness, it would be interesting to compare the cumulatedcost of partition NC with the rest of possible partitioningapproaches. Nevertheless, note that the number of ways topartition a set of |N | agents into nonempty components isgiven by the Bell number [36]
B|N | =
|N |∑s=0
1
s!
s∑j=0
(−1)s−j(s
j
)j|N |
, (37)
VOLUME 4, 2016 7
F. J. Muros et al.: A Game Theoretical Randomized Method for Large-Scale Systems Partitioning
which becomes computationally infeasible for LSS so ascomparing NC with this full set. To solve this issue, hereit is considered the subset composed of all partitions N swi
C ,which differ from NC in the fact that only one agent i ∈ Nswitches components. This set will be denoted by ΨN ,NC andits cardinality is given by
|ΨN ,NC | = |N ||NC | − |N|C|=1C |
(|N |C|=1C |+ 1
2
), (38)
where |N |C|=1C | is the number of singletons in NC . Note
that a new component could be formed assuming the agentthat switches components is not already a singleton in NC .Likewise, redundant switches between any two singletons arealso discarded from set ΨN ,NC .
Once set ΨN ,NC is introduced, some related indices will beobtained and compared with those ofNC , which can be inter-preted as a sensitivity analysis of the SVBA. In particular, theminimum, maximum, and mean cumulated costs of partitionsN swiC ∈ ΨN ,NC will be of interest. Likewise, a parameter that
computes the portion of partitionsN swiC that are improved by
NC will also be considered and symbolized by ηp, with pbeing a certain uncertainty limit, i.e., to be better than NC itis required to outperform its cost beyond a certain threshold.
Notice also that the partition with minimum cumulatedcost from ΨN ,NC , say N (1)
C , provides a suboptimal solu-tion that improves NC . Consequently, it is possible to op-timize recursively the SVBA in a greedy fashion by us-ing the minimum-cost solutions N (r)
C from the successivesets Ψ
(r−1)N ,NC , whose elements in turns admit that only one
agent switches components from partition N (r−1)C , with r ∈
N+, r > 1, i.e.,
N (r)C = arg min
N swiC
Jcum(N swiC )
s.t. N swiC ∈ Ψ
(r−1)N ,NC ,
(39)
until reaching any pre-established stopping criterion, e.g.,a maximum number of iterations rmax or a minimum per-formance improvement between iterations. This optimiza-tion represents a fine tuning of the SVBA, denoted hereby SVBA-FT. Note that, as considered in the SVBA, someadditional constraints should be included by the control de-signer in the SVBA-FT to balance the size of the components.Likewise, note that sets Ec, Ee and EeC could be modified as aconsequence of the fine tuning procedure. In any case, thesechanges improve the performance of the optimized solution,symbolized by N opt
C .
Remark 6. Given how ΨN ,NC is built, the impact on com-munication burden between two consecutive optimizationsteps is negligible. Nevertheless, when a high number ofsteps is performed, the aforementioned size constraints andalso a stopping criterion are necessary to avoid inefficientcentralized partitionings.
TABLE 1. Set ΨN ,NC related toNC described by (35)
Example 3. Let the solution of the SVBA described by (35).Following (37), the full set of different solutions for six agentsis given by B6 = 203, whereas the cardinality of ΨN ,NC isreduced to 17, according to (38). All partitions included inthat set are detailed in Table 1, where the switching agentfor each case with respect toNC is represented in blue color.Note that the cases of any agent in C1 or C2 switching to a newsingleton C4 are taking into account. Consider for instancethat N opt
C , i.e., the solution after performing the fine tuning,is given by Partition 14 in Table 1 as
Copt1 = {1, 2, 3}, Copt
2 = {4}, Copt3 = {5, 6}. (40)
Then, note that link {5, 6} should be removed from set Ee.Consequently, set Eopt
eC and the final network configuration forthe SVBA-FT would be respectively described by
EopteC = ∅, (41a)
(N optC , Eopt
C \EopteC ) =
({Copt
1 , Copt2 , Copt
3
}, EoptC), (41b)
with EoptC = N opt
C ×NoptC .
V. CASE STUDYThe proposed partitioning scheme has been implementedin the Barcelona DWN, which is managed by Aguas deBarcelona, S.A. (AGBAR). The Barcelona DWN distributesthe water supplied by the Ter and Llobregat rivers, which areregulated at their head by dams with an overall capacity of600 hm3, to the whole Barcelona metropolitan area. Besidesthe rivers, some additional underground wells also contributeto an overall flow of around 7 m3/s, which becomes potableby using four drinking water treatment plants. Given thelimits in the water flow provided by each source, there existdifferent water prices depending on water treatments andlegal extraction canons.
A. COALITIONAL CONTROL MODELControl-oriented schemes for DWNs have been widely an-alyzed in the literature [37], [38]. In particular, the controlapproach of the full Barcelona DWN discussed in [13], [14]
8 VOLUME 4, 2016
F. J. Muros et al.: A Game Theoretical Randomized Method for Large-Scale Systems Partitioning
c70PAL
c125PAL
iPalleja4
d110PAP
c110PAP
iPapiol2AGBAR
d54REL
d100CFE
c100CFE
vSJD
c100LLO
d80GAVi80
CAS85
c80GAVi80
CAS85
c70LLO
vCanyars
iRelleu
iGava4
iCastelldefels
d115CAST
c115CAST
vCanRoca
iBellamar
dPLANTAAportLL1
iSJD70
iSJD50
cRECARREGA
iSJD10
iCornella50
iCornella70
c70FLL
vZonaFranca
vCerdaTraja
vTorrassa
vRossichMaq
vCollblanc
iCollblanc
iEsplugues
vEsplugues
c130BAR
iFinestrelles200iFinestrelles176
d200BLL
c200BLL
vFinestrelles
d176BARsud
c176BARsud
c200BARsc
vBonanova
vPortola
vMinaCiutat
d200ALT
c200ALT
vBaroStLluis
c200BARnordd90CERD
c101MIR
iAltures
vPsgStJoan
c100BLLnord
d70BBE
c70BBE
iCornella100
iCornella130
iRoquetes
aMS
iMinaSeix
aPousCAST
iPousCAST
aPousB
aPousE12
iEtapBesos
iEstrella12
d130BAR
c140LLO
d125PAL
AportT
vPalleja70
vAbrera
vFontSanta
vGava100a80
vTrinitat70
vTrinitat100
vTerMontcadavTerCarmel
vTrinitat130
AportLL2
c176BARcentre
AportA
vTrinitat200
vSifoTer
vBesosMontcCerd
v70LLFLL
vTer
iSJDSpf
iSJDSub
d205FON d320FON
iPalleja1 iPalleja2
c205FON c320FON
c356FON
d175PAP
c175i135PAP
vPapiolATLL
c82PAL
d130LSE
iLaSentiu
c130LSE
aCAST8
iCastelldefels8
d145MMA
iMasJove
c145MMA
c175BVI
iStBoi
iViladecans1
iStCliment1
iBegues1
d150SBO
d175LOR
iOrioles
d184BEG
d255BEG
iBegues2
d114SCL
d190SCL
iStCliment2
d135VIL
d185VIL
iViladecans2
d369BEG
iBegues3
d450BEG
iBegues4
C175LOR C185VIL c190SCL c255BEG
c150SBO C135VIL c114SCL C184ESP
c369BEGc450BEG
iCanRoig
c195Tor
d147SCC
iStaClmCervelloc147SCC
d205CES
iCescaplina1
c205CES
d263CES
iCescaplina2
c263CES
d252CGL
d313CGL
iCanGuell2d3
d246CGYd200CGY
d268CGY
c200CGY c246CGY c252CGL
iCanGuey1d2 iCanGuey1d5 iCanGuell1
iCanGuey2
c268CGY
d361CGY
c361CGY
iCanGuey3
d374CGL
iCanGuell2d5
c374CGL c313CGL
d200BSO
iBellsoleig
c200BSO
d437VVI
c300BAR
iFinestrelles300
d320MGB
iPapiol1
c320MGB
C437VVI
iMasGuimbau1
iTibidabo
d400MGB
iMasGuimbau2
c400MGB
c475MGB
c541TIB
d255CAR
iCarmel
c255CAR
d260SGE
d328SGE
c260SGE
c328SGE
iStGenis1
iStGenis2
c250TBA
iTorreBaro2
iCerdUAB
d197BET
d215VALL
d184SMM
iStaMaMontcada
d132CMF
d200FDM
iFlorMaig
c238UAB
c200FDM
c184SMM
c197BET
c132CMF
C215VALL
c250VASAB
iVallensana1
c260VALL
c275BEV
d169CME
iTorreBaro1
c169CME
c55BAR
d120POMiMorera
iMntjcStaAmalia
iMntjcTresPins
d171SAM
d144TPI
d190TCA
c144TPI
C171SAM c190TCA
c120POM
d117MTG
c117MTG
c135SCG
d151BON
d197GUI
c151BON
c197GUI
d225GUI
c225GUI
iBonavista
iGuinardera1
iGuinardera2
c70CFE
vBesosStaColoma
c100BES
iTorreoCastell
vFinestrllEsplg
c150ALT
c176BARnord
d202CRU c202CRU
iCanRuti
d1 d2
d3 d4 d5 d6
d7 d8 d9 d10
d11
d12
d13
d14 d15
d16
d17 d18
d19 d20 d21
d22
d23 d24 d25
d26
d27
d28 d29
d30
d31
d32
d33
d34
d35
d36
d37
d38
d39
d40
d41
d42
d43
d44
d45
d46
d47 d48
d49
d50
d51
d52
d53
d54d55
d56
d57
d58
d59
d60
d61
d62
d63
d64
d65
d66d67
d68
d69
d70
d71
d72
d73
d76
d77
d78
d79
d80
d81
d82
d83
d84
d86
d85
d87
d88
x1x2
x3 x4 x5 x6
x7 x8 x9 x10
x11
x12
x13
x14
x16 x17
x18
x19 x20
x21
x22 x23 x24
x25
x26
x27 x28 x29
x31
x32x33
x30
x34
x35
x36
x37
x38
x39
x40
x41
x42
x43
x45
x46
x47
x48
x49
x50
x51x52
x54
x53
x55
x58
x56
x57
x59
x60
x61
x63
x62
u1
u2
u3 u4 u5 u6
u7
u8
u9
u10
u11
u12
u13
u14
u115
u15
u16
u17
u18 u19 u20
u21
u26
u23 u24
u30
u31u28
u34u35
u109
u111u108
u32
u38
u40
u37
u112
u81
u123
u47
u42
u39
u43
u44
u45
u52
u56
u57
u60
u62
u66
u61
u63u65u68
u69
u67
u85
u50
u51
u78
u82
u105
u88u102
u92
u94
u101
u100
u29
u110u107
u36
u33
u41
u46
u55
u86
u53
u125
u90
u76
u74
u72u71
u106
u83
u89
u80
u73
u87
u103
u97
u91
u96
u93
u98
u95
u99
u49
u22
u27
u25
u48
u58
u113
u84
u64
vConflent
v70CFE
vMontigala
iUAB
vHorta
iMasGuimbau3
iPalleja3
iVallvidrera
d300BAR
v55BAR
iCerdSabadell
iVallensana2
iBoscVilaro
d10COR
d195TOR
u59
u104
vCncpcioArenal
u70
c100BLLsc
d74
aPousE3456
iEstrella3456
u77
u75
135MGA
vAltures
v117Montigala
iMontemar
vPallejaATLL
vTerStaColoma
vSJDTot
vPousEstrella
iCerdMontflorit
vCerdanyola90
x44
vCornella100
x15
u116
u117
u118
u119
u120
u121
u122
u124
u54
u126
u79
u127
u128
u129
u114
u130
N4
N3
N2N1
N5
N6
N7
N8
N9
N10
N11
N12
N13
N14 N15
N16
N17
d75
Last Update:
23 Mar, 2011
Juan Manuel Grosso
SAC – IRI – UPC
63 STATE VARIABLES (x)
114 CONTROL ACTIONS (u)
130 FLOW ACTUATORS (u)
88 DISTURBANCES (d)
17 NODES (N)
Model Parameters Summary
ReservoirDemand
sector
Source
(Superficial or
Underground)
Pumping
stationValve
LEGEND: Infrastructures
Distribution
node
FIGURE 3. Entire model of the Barcelona DWN [13], [14]
is considered in this paper and depicted in Fig. 3. This modelcontains a total amount of 63 tanks, 114 actuators – dividedinto 75 pumps and 39 valves – and 88 sectors of consume thatrepresent the external disturbances. A graph that summarizesthe physical connections among the storage tanks and thejunction nodes is provided in Fig. 4.
In the approach proposed in this paper, as commentedbefore, an initial pre-partitioning into agents is performed dueto the node equations that appear in the system (see Fig. 3),e.g.,
FIGURE 4. Graph representing the full Barcelona DWN, where the tanks aredenoted by x, the junction nodes by N, and with the arrows representing thedirections of the flows.
.u(25)− u(27)− u(28)− u(29)− u(105)− d(15) = 0,
u(27)− u(26)− d(36) = 0,(42)
physically connect flows u(25), u(26), u(27), u(28), u(29)and u(105). Hence, the values of these flows must be simulta-neously determined. For this reason, they are assigned to thesame agent. As a consequence, the only coupling among thesubsystems is due to the inputs effect in the dynamical model.Therefore, in the case study, Aij = 0 in (1), and equivalently,ACj = 0 in (4). Considering this approach, 43 agents havebeen obtained, where the criterion of considering outflowsas disturbances has been assumed, i.e., agents control theirinflows. The explicit definition of each agent is included inAppendix A, where the identification of Agent 1 is detailedas an example.
Remark 7. Following this approach, the constraints imposedby the node equations are assigned to a given agent and arealways satisfied. This fact represents an advantage with re-spect to the partitioning performed in [13], where the agentsdo not satisfy in general the node equations and the resultingcomponents need to communicate partially following a hi-erarchical structure and generating virtual demands amongthem, i.e., they cannot work in a truly decentralized fashion.
VOLUME 4, 2016 9
F. J. Muros et al.: A Game Theoretical Randomized Method for Large-Scale Systems Partitioning
From an overall centralized viewpoint, the following equa-tions are satisfied:
with xN (k) ∈ R63, uN (k) ∈ R114 and dN (k) ∈ R88.This way, (43a) corresponds with the dynamics of the storagetanks, and (43b) describes the network static relations due tothe mass balance at each of the 17 junction nodes (see Fig. 3),with EN ∈ R17×114 and EdN ∈ R17×88 being weightingmatrices of proper dimensions.
Finally, consider the main physical constraints of the DWNgiven by the variables related to the tank volumes and manip-ulated flows, i.e., ∀k
xminN ≤ xN (k) ≤ xmax
N ,
uminN ≤ uN (k) ≤ umax
N .(44)
Remark 8. Soft constraints have been introduced to imple-ment the state constraints in (44). This fact, combined withthe pre-partitioning based on the node equations, avoid in-feasibility issues when solving the optimization problems.
Remark 9. In order to assess the impact of the links from adecision-making viewpoint, it is assumed that each agent hasaccess to overall state xN (k), and knows how its decisionsaffect the overall system. However, decisions can only becoordinated inside coalitions, i.e., even when the differentcoalitions try to optimize the overall system performance,they cannot agree upon the value of the system variables.
B. CONTROL OBJECTIVE: SYSTEM MANAGEMENTCRITERIA
The following management policies for the Barcelona DWNare considered given the knowledge of the system and theperformance objectives to be reached (see [13], [14] for de-tails):
• Minimizing drinking water production and transportcosts due to chemicals, legal canons and electricitycosts, which are expressed as
f1,i(k) = (We1α1 + We2α2(k))Tui(k), (45)
where vector α1 ∈ Rnui corresponds to water costs,vector α2(k) ∈ Rnui considers time-dependent elec-tricity costs, and matrices We1 ,We2 ∈ Rnui
×nui addthe corresponding prioritization to the aforementionedcosts within the related multi-objective optimizationproblem.
• Maintaining the stored volume around a given safetyvalue in case of emergency, which is achieved by mini-mizing
f2,i(k) = ςTi (k)Wxςi(k), (46)
withxi(k) ≥ xsaf
i − ςi(k) ≥ 0, (47)
where xsafi ∈ Rnxi is a vector of safety volume thresh-
olds in m3 (conveniently determined according to themanagement company policies related to the DWN),with ςi(k) ∈ Rnxi representing the amount of volumegoing down from the desired thresholds, and whereWx ∈ Rnxi
×nxi is a weighting matrix.
• Penalizing sudden variations of the control inputs byminimizing
f3,i(k) = ∆uTi (k)R∆u∆ui(k), (48)
where ∆ui(k) = ui(k) − ui(k − 1), and with R∆u ∈Rnui
×nui also being a weighting matrix.
Hence, the individual cost related to agent i ∈ N that isconsidered in this paper is given by
`i(xi(k),ui(k))
= f1,i(k) + f2,i(k) + f3,i(k). (49)
Finally, the aggregate cost of a certain communication com-ponent C is defined by
`C(xC(k),uC(k))
=∑i∈C
`i(xi(k),ui(k)), (50)
given that no couplings on the cost are considered inthis work.
Remark 10. Considering how (50) is built, (8) results in aquadratic programming (QP) problem. Therefore, convexityis guaranteed in the proposed approach, which allows a fastcalculation of the solution for each optimization problem anda computation of (9) in polynomial time, as required in [28].
VI. SIMULATION RESULTSThe Shapley-value-based partitioning algorithm presentedin this paper has been tested for the Barcelona DWN byusing the Matlab® solver quadprog in a 3.4 GHz IntelOcta-Core® i7-6400, 16 GB RAM computer. This way, acoalitional MPC scheme has been implemented in closedloop by considering Tsim = 24 simulation instants (one day),and with Np = 12. The numerical values of the performanceparameters are determined by a trial-and-error procedure,resulting in We1 = 0.9I, We2 = 0.5I, Wx = 10I,R∆u = 0.1I, with I being the identity matrix of suitabledimensions. Note also that there is no reference when consid-ering (46), i.e., the controller chooses the most appropriatewater volumes that satisfy the soft constraints imposed byxsafi . Likewise, the initial state is constant for all possible
coalitions and is slightly above the minimum safety level.The Shapley values for the 43 agents cannot be directly
computed due to computational issues, which are solved herevia the randomized method [28] introduced in Section III-A,considering ε = 0.2σφi , ∀i ∈ N , λ = 0.1, and Zλ/2 =1.6449, which, in order to verify (20), requires a sampleQ with q = 68 permutation vectors. According to (3), the
10 VOLUME 4, 2016
F. J. Muros et al.: A Game Theoretical Randomized Method for Large-Scale Systems Partitioning
TABLE 2. Indices ξl(N ,v) for the 20 best/worst performance links
43 agents are related to 903 possible communication links.Likewise, the redistribution of the estimated Shapley value ofthe agent-based game to the links is obtained by using (11).Notice that for the estimation of the Shapley value, 43·q coali-tions have been evaluated by means of (9). The performanceof these coalitions in terms of their cardinality is representedin Fig. 5, where it can be seen a low correlation betweencardinality and coalition performance. In other words, thekey of the partitioning performance is not related to groupas many agents as possible but to select the clusters ofcooperating agents properly.
The corresponding index ξl(N ,v) for every link l = {i, j}is represented in a color scale in Fig. 6. Likewise, the valuesfor the 20 best/worst performance links are explicitly indi-cated in Table 2.
A. SHAPLEY-VALUE-BASED PARTITIONINGAPPROACH (SVBA)
The proposed partitioning approach has been tested with thefollowing thresholds, which have been determined by a trial-and-error tuning procedure:
Lc = 1.0× 107, (51a)
γ = 2.5, ν = 2.3, ρ = 100, (51b)
Le = 1.5× 107, (52)
FIGURE 5. Cost of the coalitions needed by the randomized method in termsof their cardinality. Note that there are coalitions with few agents and properperformance – closer to the grand coalition – and vice versa, i.e., coalitionswith many agents and performance similar to the empty coalition.
FIGURE 6. Estimated redistributions of the Shapley value. Yellowest andbluest colors represent most expensive and cheapest links, respectively.It can be seen that agents 2, 3 and 38 are endpoints of the links with bestperformance, whereas agents 23 and 40 are related to the links with theworst ones.
where several requirements have been considered in the ad-justment of the aforementioned thresholds. In the first place,a cardinality constraint for any component of 0.2|N | hasbeen imposed, for our primary goal is to avoid componentslarger than one fifth of the system agents. Secondary goalswere also considered, specifically to increase the cardinalityof the resulting components so as to reduce the numberof singletons.
Notice that Ee is completely delimited by Le. Likewise,with parameters in (51b) set to zero, set Ec would alsobe completely determined by Lc. Under this premise, it ispossible to represent the cardinality of both sets Ec and Eeas a function of any threshold corresponding to either Lc orLe, which is depicted in Fig. 7. Notice that both functions aresymmetric given that in the limit case, i.e., Lc = Le = L ,it is trivially verified
|Ec(L )|+ |Ee(L )| = |E|. (53)
FIGURE 7. Cardinality of sets Ec and Ee as a function of a given threshold L ,for the particular case of not considering parameters to balance the size of theresulting components, i.e., γ = ρ = 0. The red crosses refer to the chosenthresholds, i.e., |Ec(Lc)| = 699 and |Ee(Le)| = 65.
VOLUME 4, 2016 11
F. J. Muros et al.: A Game Theoretical Randomized Method for Large-Scale Systems Partitioning
As seen in Fig. 7, set Ec for γ = ρ = 0 would be composedof 699 links. Note that these links connect all agents in N ,achieving the grand coalition, i.e., NC = {N}. For thisreason, (51b) is considered to penalize not only incominglinks in Ec(s), but also their remaining neighbors in E\Ec(s).As a result, the following 52 links have been obtained anddrawn in Fig. 8 in a color scale between green and yellow,with darkest links representing the useful ones:
Finally, the overall network after performing the parti-tioning by the SVBA would be described by (55) and (57),achieving the network configuration specified in (33).
B. PARTITIONING APPROACH FINE TUNING (SVBA-FT)
The partitioning approach has been optimized for a simula-tion of a day with average demand and disturbances, startingby finding the partition with minimum cumulated cost N (1)
Cfrom set Ψ
(0)N ,NC = ΨN ,NC , which in turns is obtained from
TABLE 3. SVBA-FT Optimization Procedure
r Switches with respect to (55) Jcum(N (r)C ) γ(r)
1 Agent 43 from C10 to C17 9.7859 ×107 –
2 Agent 5 from C1 to C2 9.7096 ×107 –
3 Agent 43 from C17 to C1 2.5892 ×107 27.44 %
4 Agent 9 from C13 to C2 2.5098 ×107 25.73 %
5 Agent 11 from C6 to C14 2.4858 ×107 25.78 %
6 Agent 22 from C4 to C10 2.4619 ×107 1.66 %
7 Agent 17 from C8 to C5 2.4429 ×107 0.90 %
solution NC described by (55). This optimization procedurehas been recursively applied considering size constraintssimilar to those of the SVBA, i.e.,
(60)where a 3-step average performance improvement has beenconsidered, taking Jcum(N (0)
C ) = Jcum(NC) = 1.0662×108,r ∈ N+, r > 3.
The results related to every iteration are detailed in Table 3,where only rstop = 7 steps have been needed to fulfill thestopping criterion described by (59), which indicates that thesolution in (55) is indeed a suitable starting point. Note that
FIGURE 8. Links belonging to Ec after applying the partitioning procedure,with their performance normalized. This way, value “0” refers to the link withbest performance, i.e., {2, 38}, and it is drawn in pure green, and value “1”is related to the link with worst performance out of the 52 links in Ec,i.e., {11, 25}, and it is represented in pure yellow.
12 VOLUME 4, 2016
F. J. Muros et al.: A Game Theoretical Randomized Method for Large-Scale Systems Partitioning
the number of partitions explored by the greedy procedurecan be easily calculated by
∑rstop−1r=0 |Ψ(r)
N ,NC | = 6143. Thedefinitive solution after applying the optimization procedureis determined by
N optC =
{ Copt1︷ ︸︸ ︷
{1, 2, 3, 27, 32, 37, 38, 42, 43},
Copt2︷ ︸︸ ︷
{4, 5, 9, 15, 16, 29},Copt3︷ ︸︸ ︷
{10, 17, 19},
Copt4︷ ︸︸ ︷
{24, 34, 36},
Copt5︷ ︸︸ ︷
{11, 12},
Copt6︷ ︸︸ ︷
{13, 14},
Copt7︷ ︸︸ ︷
{18, 26},Copt8︷ ︸︸ ︷
{21, 22},
Copt9︷︸︸︷{6} ,
Copt10︷︸︸︷{7} ,
Copt11︷︸︸︷{8} ,
Copt12︷︸︸︷{20},
Copt13︷︸︸︷{23},
Copt14︷︸︸︷{25},
Copt15︷︸︸︷{28},
Copt16︷︸︸︷{30},
Copt17︷︸︸︷{31},
Copt18︷︸︸︷{33},
Copt19︷︸︸︷{35},
Copt20︷︸︸︷{39},
Copt21︷︸︸︷{40},
Copt22︷︸︸︷{41}
}.
(61)
Finally, note that the switches performed as a result of theoptimization procedure do not imply to remove any commu-nication link in set Ee. Consequently, set EeC remains constantwith respect to (57), being the final network configurationdescribed by
(N optC , Eopt
C \EeC ), (62)
with EoptC = N opt
C ×NoptC . In any case, note that cheaper/more
expensive agents illustrated in Fig. 6 are not affected by thechanges introduced by the fine tuning.
Remark 11. Given that the proposed procedure optimizesdiscrete variables, i.e., the components, γ(r) might increasein further iterations with r. That is, there is no guaranteethat the imposed limit becomes a bound for later iterations.Nevertheless, the average way in which γ(r) is defined miti-gates this possibility.
C. COMPARISON OF THE PROPOSED PARTITIONINGSOLUTIONS WITH OTHER SCHEMES IN THELITERATURE
In this section, both the original and refined proposedpartitioning solutions will be compared with the solu-tions obtained by applying other different partitioning ap-proaches [10], [13] to the full Barcelona DWN, and also withthe centralized and decentralized configurations.
1) An alternative Barcelona DWN partitioning scheme
In [13], an alternative partitioning method of the sameBarcelona DWN model analyzed in this paper, which basi-cally consists in a graph-theory-based approach (GTBA), isproposed. Nevertheless, that work follows a different crite-rion to define the agents, and for this reason their resultsare not directly comparable with the approach proposed here.With the aim of providing a way to compare all approaches,it has been considered that each of the 43 agents definedhere belong to a component described in [13] if all its relatedvariables, i.e., states, inputs, disturbances, are contained intothis component. Under this assumption, which represents an
approximation of the partitioning in [13], most of the 43agents have been distributed into the six components in [13],with the exception of agents 1 and 2, which have beenassumed to belong to new independent components. Takingthis fact into account, the partitioning provided by the GTBAcan be modeled by
(63)where it can be seen that this solution does not respectthe size constraints imposed to the SVBA/SVBA-FT, whichrepresents an advantage for the GTBA.
2) A partitioning approach applied to other DWN
In [10], a partitioning approach based on social networkalgorithms (SNBA) is introduced and applied to the PareteDWN, located in the South of Caserta, Italy. Basically, thisapproach uses the centrality index called edge betweennessas a metric to identify the boundaries of communities [39].Let (V,L) be a directed graph that describes the direction ofthe flows, symbolized by links lf ∈ L, among any pair ofvertices {s, t} ∈ V , which are related to water entities, e.g.,tanks. The edge betweenness cB(lf) of a link lf is defined asthe number of optimal paths between vertex pairs that runalong link lf, summed over all vertex pairs, as follows [40]:
cB(lf) =∑{s,t}∈V
σ(s, t|lf)σ(s, t)
, (64)
where σ(s, t) is the number of shortest (s, t)-paths, andσ(s, t|lf) is the number of those paths passing through link lf.This way, an optimal community cluster can be definedby progressively removing edges with high value of edgebetweenness from the original graph [39]. In other words,index (64) identifies edges in a network that lie betweencommunities, which can be progressively removed leavingbehind just the communities themselves.
Note that for the Barcelona DWN case study, the directionof the flows among the water tanks and junction nodes isdescribed by graph in Fig. 4. Given that in this work a pre-partitioning into agents has been made, to properly apply thescheme proposed in [10] it is needed to map the previousgraph into a new one that depicts the direction of the flowsamong the agents. As a result, graph (V,L) represented inFig. 9, with V = N = 43 agents and L = 49 directed linkshas been obtained, where each agent comprises informationregarding several tanks and nodes following the criterionestablished in Section V-A. Once graph (V,L) is established,
VOLUME 4, 2016 13
F. J. Muros et al.: A Game Theoretical Randomized Method for Large-Scale Systems Partitioning
FIGURE 9. Graph (V,L) representing the direction of the flows among the43 agents for the full Barcelona DWN. As it can be seen, agent 2 has a strongcentrality regarding the flows.
it is possible to compute index (64) for all these 49 directedlinks. Finally, in order to obtain the partitioning solution, thelinks with higher edge betweenness have been progressivelyremoved until achieve the stopping criterion of a maximumcardinality for any component of 0.2|N |, as done in thepartitioning approach proposed in this paper. As a result,18 links have been removed, which are depicted by greendashed arrows in Fig. 9, and explicitly represented joint totheir cB(lf) in Table 4. The resulting partitioning scheme isgiven below:
N SNBAC =
{ CSNBA1︷ ︸︸ ︷
{2, 6, 7, 8, 29, 31, 33, 36, 41},CSNBA2︷ ︸︸ ︷
{1, 3, 21, 22, 23, 25, 37, 38},
CSNBA3︷ ︸︸ ︷
{15, 16, 19},
CSNBA4︷ ︸︸ ︷{4, 5},
CSNBA5︷ ︸︸ ︷
{10, 11},
CSNBA6︷ ︸︸ ︷
{13, 14},
CSNBA7︷ ︸︸ ︷
{18, 20},
CSNBA8︷ ︸︸ ︷
{24, 26},
CSNBA9︷ ︸︸ ︷
{27, 42},CSNBA10︷ ︸︸ ︷
{28, 43},
CSNBA11︷ ︸︸ ︷
{30, 32},
CSNBA12︷ ︸︸ ︷
{34, 35},
CSNBA13︷ ︸︸ ︷
{39, 40},CSNBA14︷︸︸︷{9} ,
CSNBA15︷︸︸︷{12},
CSNBA16︷︸︸︷{17}
}.
(65)Remark 12. Directed graph (V,L) = (N ,L) has nothing todo with complete undirected graph (N , E), which was usedin the partitioning algorithm proposed in Section IV. Theformer represents the direction of the physical flows amongthe agents, i.e., |L| = 49, whereas the latter assumes thatall agents are initially interconnected, i.e., |E| = 903, tolater classify these undirected communication links into setsEc and Ee regarding their control performance.
TABLE 4. Edge betweenness cB(lf) of directed links lf removed from graph inFig. 9 to find communities
All approaches have been tested by using the sensitivityanalysis introduced in Section IV-A, where the solutions,defined respectively by (55), (61), (63), (65), and also thecentralized (CEN) and decentralized (DEC) configurations,have been compared with those in their corresponding setsΨN ,NC . The cardinality of these sets (obtained by (38)) andthe related indices are illustrated in Table 5 for the same av-erage day used in the previous section, where an uncertaintylimit of p = 0.1% has been assumed in the computationof ηp. As expected, the value Jmin
cum (ΨN ,NC ) of the SVBAcorresponds with the first step of the optimization procedurein Table 3.
Note that |ΨN ,NC | provides the number of partitioningsolutions explored in the sensitivity analysis. Nevertheless,only the solutions of that set that satisfy the size constraints,i.e., maximum cardinality for any component of 0.2|N |,should be considered. This way, note that the CEN has beenincluded in this comparison even without any element inthe corresponding set ΨN ,NC trivially satisfying this sizeconstraint, which explains its best performance. Likewise, ascommented before, there are two components of the GTBAthat also do not respect the size constraints, which representan advantage to this scheme in the comparisons.
As can be seen, the values of Jcum(NC), Jmincum (ΨN ,NC ) and
Jmaxcum (ΨN ,NC ) for the SVBA/SVBA-FT improve those of
the GTBA, SNBA and DEC. That is, the proposed solutionsimprove the performance of the rest. According to ηp, it canbe checked that the SVBA outperforms a higher portion ofpartitioning solutions in corresponding sets ΨN ,NC than theGTBA and SNBA, and is in the order of the DEC, whichin any case is improved by the SVBA-FT. This finding isconsistent with Fig. 5, which already illustrated that only afew topological changes in the network can increase the per-formance substantially. Note as well that the SVBA/SVBA-FT also outperform the other methods when the differencebetween the cost of the different approaches and the cor-
14 VOLUME 4, 2016
F. J. Muros et al.: A Game Theoretical Randomized Method for Large-Scale Systems Partitioning
TABLE 6. Overview of the different methods considered
Scheme |NC | Jcum(NC)-30 days
CEN 1 3.7569 ×106
SVBA-FT 22 1.7422 ×108
SVBA 23 7.7357 ×108
SNBA [10] 16 1.3220 ×1011
GTBA [13] 8 1.1357 ×1012
DEC 43 2.3669 ×1012
responding minimum in their set of alternatives ΨN ,NC isexamined, i.e., the proposed approaches are closer to theseminimum-cost solutions than the other methods. Likewise,it is interesting to check that both the SVBA and SVBA-FThave a better performance than the mean performance of thatset, i.e., Jcum(NC) < Jµcum(ΨN ,NC ), which is also the case ofthe SNBA but does not occur in the GTBA and DEC. Finally,notice that all parameters in the optimized scheme SVBA-FTimprove those in the SVBA, as expected. All in all, theseresults indicate that both proposed approaches outperformthe rest of schemes.
D. PARTITIONING LONG-SIMULATION OVERVIEW
A comparison between the different schemes considered inthis paper is summarized in Table 6, where the cumulatedcost of a 30-day simulation scenario with demand and dis-turbances taken from real data has been calculated for eachapproach. Note that it is reasonable to test the results ina longer scenario than the one used for the design, i.e.,one day. It can be seen that both the SVBA and SVBA-FTimprove the results of the GTBA, SNBA and DEC, showingthe effectiveness of the partitioning algorithm proposed inthis paper.
Notice that the SVBA provides us with a suboptimalsolution within the set of B43 options (recall (37)), whichrepresents a suitable starting point for the fine tuning. Then,the goal of the SVBA-FT is to increase the performance bycarrying out a greedy search around SVBA. In particular,significant gains can be obtained by reducing any violation ofconstraints, which are severely penalized by soft constraints.Also, the fine tuning helps to mitigate possible deviationsintroduced by the randomized method used for the estimationof the Shapley value. In any case, note that the fact that ofsome links originally included in sets Ec /Ee could be removeddue to the SVBA-FT does not break any physical constraintsof the original DWN given that these constraints are usedto define the agents. This is not the case of other schemessuch us [10], where a fine tuning procedure could break somephysical/topological connections.
Finally, these partitioning results can be improved once thecommunication components start exchanging information,e.g., by using a coalitional control scheme [22], [24].
VII. CONCLUSIONS
In this paper, a heuristic partitioning algorithm for LSS hasbeen introduced considering that agents are able to operatein a decentralized fashion. This method ranks the commu-nication links inside a network from a control-performanceperspective. A game over agents based on the utilization ofcoalitional MPC to control the plant has been considered,and a randomized method has been used to estimate theirShapley value to deal with combinatorial explosion issues.The redistribution of the estimated Shapley value of theagents among the links has been proposed as a measure oftheir relevance in the control system and to fix/remove them.
The proposed algorithm calculates sets Ec and Ee. The linksin the former set will determine the communication compo-nents, where several mechanisms to avoid the formation oflarge clusters of cooperating agents have been considered.Once the components are established, some connectionsamong them are disconnected by using the information inthe latter set, providing a proper configuration of the overallnetwork. The resulting partitioning has also been refined bya method that optimizes it recursively, based on some costindices and size constraints. Both the original and refinedapproaches have been tested with the Barcelona DWN as acase study, providing reasonable solutions that outperformother partitioning schemes available in the literature.
.
APPENDIX A STATES, INPUTS AND DISTURBANCESRELATED TO THE AGENTS
As mentioned in Section V-A, the flows belonging to thefollowing node equations:
are physically interconnected, in particular by means of flowu(27). Then, all states that comprise incoming flows involvedin (66), which are represented in blue color, should belong tothe same agent, i.e.,
(67)where superindex + refers to the successors state. Finally,note that outflows in the states – drawn in red color – areconsidered here as disturbances.
By considering this approach, the following 43 Agentshave been obtained:
The authors would like to thank the Editor and the anony-mous reviewers for their valuable suggestions and comments.
REFERENCES[1] E. F. Camacho and C. Bordons, Model Predictive Control, 2nd ed., ser.
Advances Textbooks in Control and Signal Processing. London, England:Springer, 2003.
[2] J. M. Maestre and R. R. Negenborn, Eds., Distributed Model PredictiveControl Made Easy, ser. Intelligent Systems, Control and Automation:Science and Engineering. Dordrecht, The Netherlands: Springer, 2014,vol. 69.
[3] R. R. Negenborn, P.-J. van Overloop, T. Keviczky, and B. De Schutter,“Distributed model predictive control of irrigation canals,” Networks andHeterogeneous Media, vol. 4, no. 2, pp. 359–380, June 2009.
[4] D. D. Siljak, Decentralized Control of Complex Systems. Boston,Massachusetts, USA: Academic Press, 1991.
[5] A. Buluç, H. Meyerhenke, I. Safro, P. Sanders, and C. Schulz, “Recent ad-vances in graph partitioning,” in Algorithm Engineering. Selected Resultsand Surveys, ser. Lecture Notes in Computer Science, L. Kliemann andP. Sanders, Eds. Springer, 2016, vol. 9220, ch. 25, pp. 117–158.
[6] H. Meyerhenke, P. Sanders, and C. Schulz, “Parallel graph partitioningfor complex networks,” IEEE Transactions on Parallel and DistributedSystems, vol. 28, no. 9, pp. 2625–2638, September 2017.
[7] Y. Guo, S. Hong, H. Chafi, A. Iosup, and D. Epema, “Modeling, analysis,and experimental comparison of streaming graph-partitioning policies,”Journal of Parallel and Distributed Computing, vol. 108, pp. 106–121,October 2017.
[8] F. Rahimian, A. H. Payberah, S. Girdzijauskas, M. Jelasity, and S. Haridi,“A distributed algorithm for large-scale graph partitioning,” ACM Trans-actions on Autonomous and Adaptive Systems, vol. 10, no. 2, pp. 12:1–12:24, June 2015.
[9] R. Schneider, “A solution for the partitioning problem in partition-based-moving-horizon estimation,” IEEE Transactions on Automatic Control,vol. 62, no. 6, pp. 3076–3082, June 2017.
[10] A. Di Nardo, M. Di Natale, C. Giudicianni, D. Musmarra, G. F. San-tonastaso, and S. A., “Water distribution system clustering and partitioningbased on social network algorithms,” Procedia Engineering, vol. 119, pp.196–205, September 2015.
[11] C. Pizzuti, “A multiobjective genetic algorithm to find communities incomplex networks,” IEEE Transactions on Evolutionary Computation,vol. 16, no. 3, pp. 418–430, June 2012.
[12] K. Tsumura and H. Yamamoto, “Optimal multiple controlling nodesproblem for multi-agent systems via Alt-PageRank,” in Proceedings of the4th IFAC Workshop on Distributed Estimation and Control in NetworkedSystems, Koblenz, Germany, September 2013, pp. 433–438.
[13] C. Ocampo-Martinez, S. Bovo, and V. Puig, “Partitioning approach ori-ented to the decentralised predictive control of large-scale systems,” Jour-nal of Process Control, vol. 21, no. 5, pp. 775–786, June 2011.
[14] F. Tedesco, C. Ocampo-Martinez, A. Casavola, and V. Puig, “Centralizedand distributed command governor approaches for water supply systemsmanagement,” IEEE Transactions on Systems, Man, and Cybernetics:Systems, vol. 48, no. 4, pp. 586–595, April 2018.
[15] A. Rahman and Z. Y. Wu, “Multistep simulation-optimization modelingapproach for partitioning water distribution system into district meterareas,” Journal of Water Resources Planning and Management, vol. 144,no. 5, May 2018.
[16] J. Li, C.-C. Liu, and K. P. Schneider, “Controlled partitioning of a powernetwork considering real and reactive power balance,” IEEE Transactionson Smart Grid, vol. 1, no. 3, pp. 261–269, December 2010.
[17] A. Golshani, W. Sun, and K. Sun, “Advanced power system partitioningmethod for fast and reliable restoration: toward a self-healing power grid,”IET Generation, Transmission & Distribution, vol. 12, no. 1, pp. 42–52,January 2018.
[18] J. Anderson, Y. Chang, and A. Papachristodoulou, “Model decompositionand reduction tools for large-scale networks in systems biology,” Automat-ica, vol. 47, no. 6, pp. 1165–1174, June 2011.
[19] A. B. Kahng, J. Lienig, I. L. Markov, and J. Hu, VLSI Physical Design:From Graph Partitioning to Timing Closure. Springer, 2011.
VOLUME 4, 2016 17
F. J. Muros et al.: A Game Theoretical Randomized Method for Large-Scale Systems Partitioning
[20] Z. Zhou, S. Lin, and Y. Xi, “A dynamic network partition method forheterogenous urban traffic networks,” in Proceedings of the 15th Inter-national IEEE Conference on Intelligent Transportation Systems, Anchor-age, Alaska, USA, September 2012, pp. 820–825.
[21] F. J. Muros, J. M. Maestre, E. Algaba, C. Ocampo-Martinez, and E. F.Camacho, “An application of the Shapley value to perform system parti-tioning,” in Proceedings of the 33rd American Control Conference (ACC2015), Chicago, IL, USA, July 2015, pp. 2143–2148.
[22] F. J. Muros, J. M. Maestre, E. Algaba, T. Alamo, and E. F. Camacho,“Networked control design for coalitional schemes using game-theoreticmethods,” Automatica, vol. 78, pp. 320–332, April 2017.
[23] F. J. Muros, J. M. Maestre, C. Ocampo-Martinez, E. Algaba, and E. F.Camacho, “Partitioning of large-scale systems using game-theoretic coali-tional methods,” in Proceedings of the 16th European Control Conference(ECC 2018), Limassol, Cyprus, June 2018, pp. 2517–2522.
[24] F. Fele, J. M. Maestre, and E. F. Camacho, “Coalitional control: Coopera-tive game theory and control,” IEEE Control Systems Magazine, vol. 37,no. 1, pp. 53–69, February 2017.
[25] D. Bauso and G. Notarstefano, “Distributed n-player approachability andconsensus in coalitional games,” IEEE Transactions on Automatic Control,vol. 60, no. 11, pp. 3107–3112, November 2015.
[26] R. Olfati-Saber and R. M. Murray, “Consensus problems in networks ofagents with switching topology and time-delays,” IEEE Transactions onAutomatic Control, vol. 49, no. 9, pp. 1520–1533, September 2004.
[27] L. S. Shapley, “A value for n-person games,” in Contributions to theTheory of Games II. Annals of Mathematics Studies, vol. 28, H. W. Kuhnand A. W. Tucker, Eds. Princeton, New Jersey, USA: Princeton UniversityPress, 1953, pp. 307–317.
[28] J. Castro, D. Gómez, and J. Tejada, “Polynomial calculation of the Shapleyvalue based on sampling,” Computers & Operations Research, vol. 36,no. 5, pp. 1726–1730, May 2009.
[29] H. Ishii and R. Tempo, “Distributed randomized algorithms for the PageR-ank computation,” IEEE Transactions on Automatic Control, vol. 55, no. 9,pp. 1987–2002, February 2010.
[30] R. A. DeVore and V. N. Temlyakov, “Some remarks on greedy algorithms,”Advances in Computational Mathematics, vol. 5, no. 1, pp. 173–187,December 1996.
[31] J. Qin, Y. Chow, J. Yang, and R. Rajagopal, “Distributed online modifiedgreedy algorithm for networked storage operation under uncertainty,”IEEE Transactions on Smart Grid, vol. 7, no. 2, pp. 1106–1118, March2016.
[32] J. M. Maestre and H. Ishii, “A PageRank based coalitional controlscheme,” International Journal of Control, Systems and Automation. Inpress, 2017.
[33] R. R. Negenborn and J. M. Maestre, “Distributed model predictive control:An overview and roadmap of future research opportunities,” IEEE ControlSystems, vol. 34, no. 4, pp. 87–97, August 2014.
[34] R. L. Graham and P. Hell, “On the history of the minimum spanning treeproblem,” Annals of the History of Computing, vol. 7, no. 1, pp. 43–57,January-March 1985.
[35] J. Li, C.-C. Liu, and K. P. Schneider, “Distribution system restorationwith microgrids using spanning tree search,” IEEE Transactions on PowerSystems, vol. 29, no. 6, pp. 3021–3029, November 2014.
[36] H. W. Becker and J. Riordan, “The arithmetic of Bell and Stirling num-bers,” American Journal of Mathematics, vol. 70, no. 2, pp. 385–394, April1948.
[37] M. Brdys and B. Ulanicki, Operational Control of Water Systems: Struc-tures Algorithms and Applications. UK: Prentice Hall International, 1994.
[38] A. K. Sampathirao, P. Sopasakis, A. Bemporad, and P. P. Patrinos, “GPU-accelerated stochastic predictive control of drinking water networks,”IEEE Transactions on Control Systems Technology, vol. 26, no. 2, pp.551–562, March 2018.
[39] M. Girvan and M. E. J. Newman, “Community structure in social andbiological networks,” Proceedings of the National Academy of Sciences ofthe United States of America, vol. 99, no. 12, pp. 7281–7826, June 2002.
[40] U. Brandes, “On variants of shortest-path betweenness centrality and theirgeneric computation,” Social Networks, vol. 30, no. 2, pp. 136–145, May2008.
FRANCISCO JAVIER MUROS received theM.S. degree in Telecommunications Engineeringfrom the University of Seville in 2004, and thePh.D. on automation, robotics and telematics,summa cum laude, by the same university in 2017.Since 2005, he has acquired a wide experiencein several works related to the electricity sector.He received a Master’s degree in Design, Con-struction and Maintenance of High Voltage Elec-trical Transmission Infrastructure from the Comil-
las Pontifical University, Madrid in 2014, being between 2015 and 2016Coordinator of this Master. He performed a Ph.D.-stay in Maastricht, TheNetherlands, in 2016. At present he is with the Department of Systemsand Automation Engineering at the University of Seville. He is author andco-author of more than twenty publications including journal papers, bookchapters, and conference papers. He has participated in the European UnionProject DYMASOS and in several MINECO-Spain Projects. His researchinterests focus on cooperative and noncooperative game theory, coalitionaland distributed control.
JOSÉ MARÍA MAESTRE (M’14–SM’18) re-ceived the joint B.Eng. and M.Eng. degree intelecommunication engineering from the Univer-sity of Seville in 2005, and the Ph.D. on automa-tion and robotics in 2010 by the same univer-sity, where he also works as associate professor.He has also worked in LTH at Lund University(guest researcher), TU Delft (postdoc), Universityof Cadiz (assistant professor), and Tokyo Instituteof Technology (visiting professor). Besides his
Ph.D., which was awarded with the extraordinary prize of the Universityof Seville, he has a master degree in smart homes and intelligent buildingsfrom the Universidad Politécnica de Madrid, a master in telecommunicationseconomics from the Universidad Nacional de Educación a Distancia, anda master in economics and development from the University of Seville.His main research interests are the control of distributed systems and theintegration of service robots in the smart home. He has authored and coau-thored more than one hundred conference and journal papers regarding thesetopics. He is also editor of the books Service Robotics within the DigitalHome: Applications and Future Prospects (Springer, 2011), DistributedModel Predictive Control Made Easy (Springer, 2014), and Domótica paraIngenieros (Paraninfo, 2015). Finally, he is one of the founders of thetechnological firms Idener and Eskesso.
CARLOS OCAMPO-MARTINEZ (M’11–SM’13) received his Electronics Engineering de-gree and MSc. degree in Industrial Automationfrom the National University of Colombia, Cam-pus Manizales in 2001 and 2003, respectively. In2007, he received the Ph.D. degree in ControlEngineering from the Technical University ofCatalonia - BarcelonaTech (Barcelona, Spain).Since 2011, he is with the Universitat Politèc-nica de Catalunya, Automatic Control Department
(ESAII), currently as Associate Professor in automatic control and modelpredictive control. Between 2014 and 2018, he was also Deputy Director ofthe Institut de Robòtica i Informàtica Industrial. His main research interestsinclude constrained model predictive control, large-scale systems manage-ment (partitioning and noncentralized control), and industrial applications(mainly related to the key scopes of water and energy).
18 VOLUME 4, 2016
F. J. Muros et al.: A Game Theoretical Randomized Method for Large-Scale Systems Partitioning
ENCARNACIÓN ALGABA is Associate Profes-sor in the Department of Applied MathematicsII of the Advanced Engineering School at theUniversity of Seville and member of the Math-ematics Research Institute at the University ofSeville (IMUS). She received her Ph.D. degreein Mathematical Sciences from the University ofSeville. Her main research line is about Coop-erative Games on Combinatorial Structures withrelevant contributions in the area. She is one of
the Spanish representatives in the SING (Spain, Italy, Netherlands GameTheory) Committee from 2012 and member of the Scientific Committee inSING 10 (Cracow, Poland 2014), European meeting on Game Theory SING11-GTM 2015 (St. Petersburg, Russia 2015), SING 12 (Odense, Denmark2016), SING 13 (Paris, France 2017) and SING 14 (Bayreuth, Germany2018). She has acted as evaluator of projects at national and Europeanlevel and she belongs to the national expert evaluators board from 2014.She has organized diverse courses, streams in International Conferencesand workshops on Game Theory. She has been invited speaker and guestresearcher in various universities, and has given invited Summer Coursesand numerous talks in International Conferences.
EDUARDO F. CAMACHO (M’84–SM’00–F’11)received the Ph.D. degree in electrical engineeringfrom the University of Seville, Seville, Spain.
He is now a Full Professor with the Departmentof System and Automation Engineering, Univer-sity of Seville. He is author of Model PredictiveControl in the Process Industry (1995), AdvancedControl of Solar Plants (1997), Model PredictiveControl (1999), (Springer, 2004, 2nd edition),Control e Instrumentación de Procesos Químicos
(Ed. Síntesis), Control of Dead-time Processes (Springer, 2007) and Controlof Solar Systems (Springer, 2011). He has served on various IFAC technicalcommittees and chaired the IFAC publication Committee from 2002-2005.He was the president of the European Control Association (2005-2007)and chaired the IEEE/CSS International Affairs Committee (2003-2006),Chair of the IFAC Policy Committee and a member of the IEEE/CSSBoard of Governors. He has acted as evaluator of projects at national andEuropean level and was appointed Manager of the Advanced ProductionTechnology Program of the Spanish National R&D Program (1996-2000).He was one of the Spanish representatives on the Program Committee ofthe Growth Research program and expert for the Program Committee of theNMP research priority of the European Union. He has carried out reviewsand editorial work for various technical journals and many conferences. Atpresent he is one of the Editors of the IFAC Journal, Control EngineeringPractice, Editor-at-Large of the European Journal of Control and SubjectEditor of Optimal Control: Applications and Methods.
Dr. Camacho is an IEEE and IFAC Fellow. He was Publication Chair forthe IFAC World Congress 2002, General Chair of the joint 44th IEEE CDC-ECC 2005, and Co-General Chair of the joint 50th IEEE CDC-ECC 2011.