I Introduction
Gametheoretical control has drawn considerable attention in recent years due to its widespread applications. Some representative works include: (i) consensus/synchronization of multiagent systems [1]; (ii) distributed optimization [2, 3]; (iii) control in wireless networks; (iv) optimization in energy [4] and transportation networks [5, 6], just to name a few.
The content of gametheoretical control is using game theory to solve control problems in interacting setting, such as multiagent systems [2, 3]. Addressing such issues via game theory needs two steps. The first step is to view the agent as an intelligent rational decisionmaker in a game with defining a set of available actions and utility function for every player. The second step is to specify a learning rule for the designed game so that the agents can reach a desirable situation, e.g., a Nash equilibrium. Therefore, there are two basic tasks in gametheoretic control: utility design and learning rule design [7]. Utility functions describe components’ incentives, and learning rules mean how each player processes its available information to formulate a decision.
Compared with traditional methods, the advantage of gametheoretical control is that it provides a modularized design architecture, i.e. we can design utility functions and learning rules separately [8]. The separation is described as an hourglass architecture in [7], which is shown in Fig. 1.
When designing games, one idea is to make sure that the designed game falls under some special category games, such as potential games [9]. One advantage of designing the game as potential game is that there are a variety of learning rules which lead to a Nash equilibrium, e.g. myopic best response, logliner learning, and fictitious play [10][12]. Several papers have devoted to potential game based design in distributed control [13][15]. Other game design methods include wonderful life utility design [16], Shapley value utility design [17], congestion game based design [18], etc. However most of the above works provide no systematic methods on designing local information based utility functions. Here local information means that players can only get part of other players’ information when they play the designed game, such as networked game.
This paper focuses on providing a systematic method for designing finite potential game using local information. As far as we know, the most relevant works are [2] and [19]. But our work is different to theirs. [2] provided a systematic methodology for designing potential games with continuous action sets, where the utility functions are local information based. It showed that for any given system level objective function, there exists at least one method to design the local information based utility functions [2]. However, when we turn to games with finite action sets, the existence is not guaranteed. As for [19], it presented a necessary and sufficient condition for the existence of local information based utility functions. But no systematic method is provided for designing local information based utility functions. Furthermore, the learning rule used in [19] is better reply. Using better reply, local information based potential game can converge to a Nash equilibrium, but may not a maximum point of the system objective function.
The contributions of this paper are threefold: (i) A necessary and sufficient condition for the existence of local information based utility functions is obtained, which can be verified by checking whether the corresponding linear equations have solutions. (ii) A method for designing finite potential game using local information is presented when the linear equations have a solution. (iii) An example on consensus problem is provided to demonstrate the effectiveness of the design procedure.
The rest of this paper is organized as follows: Section II provides some preliminaries, including semitensor product (STP) of matrices, game theory, and problem description. Section III considers the design of potential game using localbased information. Section IV considers application of the design method to consensus problem. A brief conclusion is given in Section V.
Notations: is denoted by the Euclidean space of all real
vectors.
is the set of real matrices. . is thedimensional identity matrix.
is thedimensional zero matrix.
. is denoted by the th column of the identity matrix . () is the set of columns (rows) of . The transposition of matrix is denoted by is the subspace spanned by .Ii Preliminaries
Iia Semitensor Product of Matrices
The basic tool used in this paper is STP of matrices. We give a brief survey on STP of matrices. Please refer to [20] for more details.
Definition II.1
[20] Suppose , , and be the least common multiple of and . The STP of and is defined by
where is the Kronecker product.
Assume . By identifying we call the vector form of integer . A function is called a mixvalued pseudological function.
Definition II.2
[20] Let be a mixvalued pseudological function. Then there exists a unique row vector , such that
is called the structure vector of , and .
IiB Potential Game
A finite noncooperative game is a triple , where is the set of players, is the set of strategies of player for every , and is the utility function of player , with being the strategy profile of the game. Let be the set of partial strategy profiles other than player . Denote by set of form finite games with , , .
Using the vector expression of strategies, the utility function can be expressed as
where , and is called the structure vector of , .
The concept of potential game was firstly proposed by Rosenthal [21], whose definition is as follows:
Definition II.3
A finite game is a potential game if there exists a function , such that for every player and every
where is called the potential function of .
The following Lemma is obvious according to Definition II.3.
Lemma II.4
[22] A finite game is potential if and only if there exist functions such that for every
(1) 
where is the potential function, and .
IiC Problem Setup
Consider a multiplayer game played on the network, which is called networked game (NG). In fact every player can only obtain its neighbours’ information when playing game . The neighbors of player is denoted by , which is defined as follows,
Suppose a system level objective function is given, where . The system wants to optimize the objective function, while all players can only obtain its neighbour s information. The optimization problem can be described as
where
To solve this optimization problem, the idea is to design a potential game in which the utility function of every player only depends on its neighbours’ information, and the potential function of the designed game is
. Then using proper learning algorithm, such as logit learning
[10], players’ behavior converges to a strategy profile that maximize the objective function.Iii Potential Game Design Using Localbased Information
Iiia Utility Design Using Local Information
We consider the design of potential game using local information in this subsection. Before designing the game, an operator , called the drawing matrix, is necessary, where is a group of players in the game . Set
where
The drawing matrix is used to “draw” the strategies of players in from [19]:
where is the strategy of player . Particularly, where is the set of players except player .
Theorem III.1
Consider a utilityadjustable networked game with objective function
Then local information based utility function can be designed if and only if all the following equations have a solution
(2) 
where , , , , , , and , .
Moreover if the solution exists, the local information based utility function of player is
(3) 
Proof: If the utility function of the networked game is local informationbased, then we have
Rewrite (1) into vector form, and substituting (3) into (1) yields
where and are the structure vectors of and , respectively.
Equations (2) are called utility design equations, and are called utility design matrices.
IiiB Solving Utility Design Equations
In this subsection we first explore some properties of utility design equations. Then we design an algorithm to calculate solutions of utility design equations equations when the solutions exist.
Set
where
Denote by Let
Theorem III.2
Let be the utility design matrix of player in (2). Then
(5) 
In other words, every column of is a solution of And
where , and
Proof: We omit the detailed proof due to the space limitation.
Corollary III.3
Consider a utilityadjustable networked game with objective function
Then the following four statements are equivalent.
 (i)

The local information based utility functions can be designed.
 (ii)

The following equations have a solution
 (iii)

Suppose is defined in (2). Then,
(6)  (iv)

Proof: (i)(ii): It is obvious using Theorem III.1.
(ii)(iii): From linear algebra we know that condition (6) is the necessary and sufficient condition for equation (2) to have solutions.
(iii)(iv): From Theorem III.2 one sees that
Combining equation (6) we have the following result
which is equivalent to statement (iv). It is obvious that (iv) implies (iii).
In the following we design an algorithm to calculate the local information based utility function for each player when equation (2) has solutions.
Algorithm III.4

Construct and for each

Determine as
(7) 
Define as the subvector of the first elements of . Using (3), the general form of are calculated as follows
where is an arbitrary vector in .
Using Algorithm III.4, the local information based utility function of player is
Remark III.5

The computation complexity of Algorithm III.4 is mainly dependent on the caculation of (7), where the dimension of is It shows that the smaller the number of the neighbors is, the lower the computational complexity is. Further investigation for reducing the computation complexity of Algorithm III.4 is necessary.

The method proposed in this paper can also be applied to design statebased potential game [23] with local information based utilities.
IiiC Selecting Proper Learning Rule
After designing the utility function, another thing to be considered is selecting proper learning algorithm in potential games. The learning algorithm should ensure that players’ behavior converges to a stategy profile that maximize the objective function using local information. The logit learning, which is shown as follows, satisfies the above demands:

At time , the updating player selects a strategy according to the following probability
where is exploration parameter.

All other players repeat their previous actions, i.e., , where
Theorem III.6
[10] Consider a repeated potential game with potential function where all player use logit learning. The stationary distribution is
(8) 
where
is the set of probability distributions over
Moreover, the set of stochastically stable states is equal to the set of maximizers of , where a state is called stochastically stable if .Remark III.7
Since for local informationbased potential game we have . Using the above logit learning rule, the local informationbased potential game will maximize the potential function with arbitrarily high probability when is sufficiently large. For general potential game if only the local information is allowed to use, the result is not assured. Here general potential game is potential game whose utility functions are not local information based.
Iv Application: Consensus Problem
In this section we deal with consensus problem using the above results.
Consider a multiagent system with a system level objective function . The goal of the multiagent system is to maximum the objective function. Due to its mobility limitations, agent can only select strategies from a restricted strategy set. For example, a robot in 2D environment can only move to a position within a radius of its current location. Agent can communicate with its neighbors, which means that at each period the agent can only observe its neighbors’ strategies.
A question is: If the player can only select strategies from a restricted strategy set, can the local informationbased potential game maximize the objective function when is sufficiently large using logit learning. Unfortunately, the answer is no. But binary restrictive logit learning, which was introduced in [1], can accomplish the aim. Denote by the set of strategies available to player at time . The binary restrictive logit learning is as follows:

At each period , player is chosen with probability and allowed to update its strategy;

At time , the updating player selects a trial strategy from according to the following probability
where .

After selecting a trial strategy , player chooses its strategy as follows:
(9) where is exploration parameter, and

All other players repeat their previous actions, i.e., , where
If the restricted strategy sets satisfy the following conditions

Reversibility:

Feasibility: For any two strategies there exists a series of strategies such that
Then using the above binary restrictive logit learning potential game with restricted strategy sets will maximize its potential function with arbitrarily high probability when is sufficiently large [1].
Example IV.1
Consider a multiagent system with three agents . The system level goal is that all agents gather at point in 2D environment, such as a room. Each agent has a strategy set . Due to its mobility limitations, every agent can only move to a position within a radius of its current location, and there is an obstacle in , as shown in Fig. 2. The communication graph is timeinvariant, which is shown in red line. Agent can only communicate with agent . Agent can communicate with agent and . Agent can only communicate with agent .
a). Vector expression of strategies
Identify
The system level objective function can be described as
(10) 
where . Using the vector expression of strategies, the structure vector of can be calculated as
b). Local information based utility design
Set
where , and
We can verify that the following equation has solutions
According to Theorem III.1, the local information based utility function is designable. The designed local information based utility function has the following form:
where , and
c). Simulation results
The initial configuration of all agents is shown in Fig. 2. Let , and Using binary restrictive logit learning with parameter , we have the following simulation results. As , local information based potential game converges to an equilibrium point which maximizes the potential function. Then all agents agree to stay at destination forever, which are shown in Fig. 3Fig.5. Here
V Conclusion
This paper investigates the design of potential game using local information. We firstly present a necessary and sufficient condition for the existence of local information based utility functions, which can be verified by checking whether a series of linear equations have a solution. Local information based utility functions can be designed using the solutions when the linear equations have solutions. Then a consensus problem is used to demonstrate the effectiveness of the proposed design approach.
Open and interesting questions for further investigations include: If the utility functions cannot be designed as a potential game using local information, can we design a nearpotential game [24]? To maximize the system level objective function, how “near” the designed game should be for a given learning rule?
References
 [1] J. R. Marden, G. Arslan, and J. S. Shamma, “Cooperative control and potential games,” IEEE Trans. on Systems, Man, and Cybernetics, Part B: Cybernetics, vol. 39, no. 6, pp. 13931407, 2009.
 [2] N. Li and J. R. Marden, “Designing games for distributed optimization,” IEEE Journal of Selected Topics in Signal Processing, vol. 7, no. 2, pp. 230242, 2013.
 [3] B. Yang and M. Johansson, “Distributed optimization and games: A tutorial overview,” Networked Control Systems, vol. 406, pp. 109148, 2010.
 [4] W. Saad, Z. Han, H. Poor, and T. Basar, “Gametheoretic methods for the smart grid: an overview of microgrid systems, demandside management, and smart grid communications,” IEEE Signal Process. Mag., vol. 29, pp. 86105, 2012.
 [5] P. N. Brown and J. R. Marden, “Studies on robust social influence mechanisms: Incentives for efficient network routing in uncertain settings,” IEEE Control Systems, vol. 37, no. 1, pp. 98115, 2017.
 [6] X. Wang, N. Xiao, T. Wongpiromsarn, L. Xie, E. Frazzoli, and D. Rus, “Distributed consensus in noncooperative congestion games: an application to road pricing,” in Proc. 10th IEEE Int. Conf. Contr. Aut., Hangzhou, China, 16681673, 2013.
 [7] R. Gopalakrishnan, J. R. Marden, and A. Wierman, “An architectural view of game theoretic control,” ACM SIGMETRICS Performance Evaluation Review, vol. 38, no. 3, pp. 3136, 2011.
 [8] J. R. Marden and J. S. Shamma, “Game theory and distributed control,” Handbook of Game Theory with Economic Applications, vol. 4, pp. 861899, 2015.
 [9] D. Monderer and L.S. Shapley, “Potential Games,” Games and Economic Behavior, vol. 14, pp. 124143, 1996.
 [10] C. Al sFerrer and N. Netzer, “The logitresponse dynamics,” Games and Economic Behavior, vol. 68, no. 2, pp. 413427, 2010.
 [11] H. P. Young, Strategic Learning and Its Limits. Oxford, U.K.: Oxford Univ. Press, 2004.
 [12] J. S. Shamma and G. Arslan, “Dynamic fictitious play, dynamic gradient play, distributed convergence to Nash equilibria,” IEEE Trans. Autom. Control, vol. 50, no. 3, pp. 312327, 2005.
 [13] E. CamposNanez, A. Garcia, and C. Li, “A gametheoretic approach to efficient power management in sensor networks,” Operations Research, vol. 56, no. 3, pp. 552561, 2008.
 [14] T. Heikkinen, “A potential game approach to distributed power control and scheduling,” Computer Networks, vol. 50, 22952311, 2006.
 [15] J. R. Marden and A. Wierman, “Overcoming limitations of gametheoretic distributed control,” in Proc. 48th IEEE Conf. Decision Control, pp. 64666471, 2009.
 [16] J. R. Marden and M. Effros, “The price of selfiness in network coding,” IEEE Trans. Inf. Theory, vol. 58, no. 4, pp. 23492361, 2012.
 [17] E. Anshelevich, A. Dasgupta, J.Kleinberg, E. Tardos, T. Wexler, and T. Roughgarden, “The price of stability for network design with fair cost allocation,” SIAM J. Comput., vol. 38, no. 4, pp. 16021623, 2008.
 [18] Y. Hao, S. Pan, Y. Qiao, and D. Cheng, “Cooperative control via congestion game approach,” arXiv preprint, arXiv: 1712.02504, 2017.
 [19] T. Liu, J. Wang, and D. Cheng, “Game theoretic control of multiagent systems,” arXiv preprint, arXiv: 1608.00192, 2016.
 [20] D. Cheng, H. Qi, and Y. Zhao, An Introduction to Semitensor Product of Matrices and Its Applications, World Scientific, Singapo, 2012.
 [21] R. W. Rosenthal, “A class of games possessing purestrategy Nash equilibria,” Int. J. Game Theory, vol. 2, pp. 6567, 1973.
 [22] D. Cheng, “On finite potential games,” Automatica, vol. 50, no. 7, pp. 17931801, 2014.
 [23] J. R. Marden, “State based potential games,” Automatica, vol. 48, no. 12, pp. 30753088, 2012.
 [24] O. Candogan, A. Ozdaglar, and P. A. Parrilo, “Dynamics in near potential games,” Games and Economic Behavior, vol. 82, pp. 6690, 2013.
Comments
There are no comments yet.