Download as pdf or txt
Download as pdf or txt
You are on page 1of 7

Ather Gattami

Research Description

Control of Networked and Complex Systems


Ather Gattami

Purpose and Aims

The rapidly increasing use of wireless communications in control of/over large networks demands a deep understanding of how to systematically design a control system that takes into account information constraints in terms of limited communication. The project goal is to develop an information theory over graphs that integrate elements and methods from information theory and control theory using a novel approach. The approach is based on recent important progress in mathematical statistics and optimization considering graphical models and semidenite programming. The results will help us to understand and tackle long-standing problems with many important applications that consider control with limited information. Applications include distributed control and control with limited communication capacity.

Survey of the Field

The fundamental problem of systematically constructing controllers to optimize a given objective and maintain stability of a given network subject to information constraints has probably been the largest research area in the control community in the last decade. To illustrate the problem of control under information constraints, we will consider the so called Cocktail Party problem in a wireless communications network. In a cocktail party, there is a number of people that communicate with each other in groups in a room. Every communicating group could be distorted by other groups. A natural way for a group to attack this problem is to raise their voice. But that would imply an increased distortion on their neighbouring groups. This process will lead to an increasing level of noise in the room. Similarly, in a communications network, we have groups of sources that try to send messages to receivers. The senders distort each other, and there is an incentive for each sender to increase the power of its signal to the receiver(s). Of course, there are limitations on the power of the signals to be sent. The question is: What is the maximum capacity that can be achieved for every communication link, and how do we control the information ow to achieve this capacity? Here, there is a strong need for a generalized channel capacity theorem, in the spirit of the simple channel capacity results by Shannon [17].

Ather Gattami

Research Description

Figure 1: Distributed control over networks. Every region (marked with a circle in the graph to the left) is controlled by a block of a structured controller (right gure). Other applications for this type of research are Internet congestion control, power networks, vehicle formation control, sensor/actuator networks, control of supply chains, automated highway, quantum physics and systems biology. Control with information structures imposed on the decision maker(s) have been very challenging for decision theory researchers. Even in the simple linear quadratic static decision problem, it has been shown that complex nonlinear decisions could outperform any given linear decision (see [21]). Important progress was made for the stochastic static team decision problems in [12] and [14]. New information structures were explored in [8] for the stochastic linear quadratic nite horizon. These were recently explored in [2] and [16] to show that the constrained linear optimal decision problem, for innite horizon linear quadratic control, can be posed as an innite dimensional convex optimization problem, given that the system considered is stable. The distributed stochastic linear quadratic team problem was revisited in [6], which generalizes previous results for tractability of the dynamic team decision problem with information constraints. An analog deterministic version of the stochastic team decision problem was solved in [6], which also showed that for the nite and innite horizon linear quadratic H control problem with bounds on the information propagation, the optimal decision is linear and can be found by solving a linear matrix inequality. In [15], the stationary state feedback stochastic linear quadratic control problem was solved using state-space formulation and covariance constraints, under the condition that all the subsystems have a common past. With common past, we mean that all subsystems have information about the global state from some time step in the past. The problem was posed as a nite dimensional convex optimization problem. The stationary output feedback version was solved in [6]. Other approaches explore homogeneous systems on graphs ([1], [5]). Heterogeneous systems over graphs were considered using approximate methods in [11] and [9]. The role of channel capacity constraints in control was discussed in [18], [19]. The work of [6] gave a complete solution to the distributed control problems under the condition that information propagates along communications links faster than the dynamics. That is, if system S aects system P dynamically at some point, then all information accessible to S should have been communicated to P from that point of time and back. 2

Ather Gattami

Research Description

w coding x decoding

0
u0

+
u0 + w

y = u1

Figure 2: Coding-decoding diagram over a Gaussian channel. Even for Witsenhausens simple example, nding the optimal controller has been a longstanding open problem and there are still published papers that compete in nding the best (nonlinear) controller. The aim of the this project is to examine general dynamic interconnections without any restrictions on the rate of information propagation.

2.1

Benchmark Problem: The Witsenhausen Counterexample

In his celebrated paper, Witsenhausen [21] demonstrated the diculty of control of dynamical systems subject to limited information by considering a stochastic two-stage control problem that can be stated as the following static team problem1 :
0 ,1

min E

ku2 + (x + u0 u1 )2 0 (1)

subject to u0 = 0 (x), u1 = 1 (x + u0 + w) x N (0, X), w N (0, W )

Here, we have two decision makers, one corresponding to u0 , and the other to u1 . Witsenhausen showed that the optimal decisions 0 and 1 are not linear because of the coding incentive of u0 . Decision maker 1 measures x + u0 + w, and hence, its measurement is aected by u0 . Decision maker 0 tries to encode information about x in its decision, which makes the optimal strategy complex. Witsenhausen found nonlinear decision functions 0 and 1 that outperformed any linear decisions. The problem above is actually an information theoretic problem. To see this, make the following modications. The appearance of u0 in the cost function is simply to prevent large values of u2 . We can take it away from the cost function and instead add the constraint 1
1

x N (m, X) means that x is a Gaussian variable with mean E x = m and covariance E xxT = X.

Ather Gattami

Research Description

OUTPUTS

STATES
STATES

... ... ...

...
STATES

OUTPUTS

STATES

OUTPUTS

(a)

(b)

Figure 3: (a) Tree structure of a Markov chain. (b) Graph structure with loops. E u2 P in the optimization problem to obtain: 0
0 ,1

min E (x u1 )2 (2) x N (0, X), w N (0, W )

subject to u0 = 0 (x), E u2 P, u1 = 1 (u0 + w) 0

The modied problem is exactly the Gaussian channel coding/decoding problem (see Figure 2)!

2.2

Distributed Filtering and Belief Propagation

To ease the presentation of the mathematical problems and the ideas of belief propagation, we will (dually) consider distribute ltering problems instead of distributed control problems. Consider the problem of N sensors distributed over a graph. Every sensor can measure a part of a global state. Mathematically speaking, we have a linear system given by X(k + 1) = AX(k) + BW (k) Yi (k) = Ci X(k) + Vi (k), i = 1, ..., N. (3)

where W (k) is white noise. Sensor i on the graph has access to the measurement Yi (k). Every sensor i can communicate with a limited number of nodes on the graph, which are described by the set Ji . Now the question is, how should information be exchanged to minimize the information ow on the graph and where sensor is estimate of the state Xi (k) converge to X(k)? The dierence between the centralized tlering problem with N = 1 graph) and the distributed with N 2 can be explained by graphical models. The centralized ltering problem can be described by the graphical model in Figure 3(a). It simply forms a Markov chain with an output at every time step. The distributed ltering problem (N 2) becomes more complex. In general, we no 4

Ather Gattami

Research Description

4 1

3 2

2 3

1 4

(a)

(b)

Figure 4: (a) A line of soldiers exchange information to their neighbours about the number of soldiers behind and ahead. (b) A swarm of soldiers with loops in the interconnection. The algorithm in (a) cannot be used, since it is not clear who is in front and who is behind. longer have a nice tree structure. Instead we have parallel lters running that possess dierent information about the state X and aect each other through information exchange, forming loops on the graph. For the case of three nodes, we could get a graph as in Figure 3(b), where the interconnected sensors are color-coded. The diculty of graphical models with loops can be explained as follows. Consider a line of soldiers with task that each soldier should know the total number of soldiers in the line. This can be done eciently using a message-passing algorithm, an instant of belief propagation, where each soldier communicates only with adjacent soldiers in the line. The front soldier in the line says the number one to the soldier behind him, the next soldier in the line says the number two, and so on (see Figure 4(a)). The same is done in the opposite direction. Then, every soldier can nd the global number of soldiers by simply adding the numbers said to him by the soldiers in front of and behind him. This trick does not work when the soldiers form a swarm (see Figure 4(b)). In graphical models, loops in graphs give complex non-convex problems. Recently, ecient semidenite programming relaxations [13] were developed and proved to be very useful in various areas of applications. These tools, combined with ideas considering belief propagation [4], where used in [20] to handle graphs with loops in a very ecient way. Since distributed control and ltering can be modeled as graphical models, the new ideas from [20] will be used to tackle these problems.

Signicance

Successful research in this project has a great impact on many disciplines such as control theory, information theory, and statistics. It is anticipated that the ideas of this project increase the understanding of the interplay between control and information. The suggested methods will also open new research avenues and give great opportunities for dierent disciplines to interact. 5

Ather Gattami

Research Description

The research has many important applications in the industry such as Communications networks, Internet congestion control, power networks, vehicle formation control, sensor/actuator networks, control of supply chains, automated highway, quantum physics and systems biology.

Preliminary Results

Consider the Witsenhausen counterexample that was discussed in Section 2.1. As was discussed earlier, it turns out that it has a very interesting interpretation as a somehow generalized communications problem. To illustrate this interesting relationship, we considered a slightly modied version of it. The following solution to the modied example by Witsenhausen (2) shows the power of the notion of mutual information (and therefore entropy) in distributed control problems. Using a basic property of mutual information, we obtain the inequality2 P 1 P 1 log2 ( ) I(x; u1 ) log2 (1 + ). 2 2 E(x u1 ) 2 N Note that the second inequality above is the Shannon channel capacity. Monotonicity of the log-function gives the inequality PN . E(x u1 )2 P +N Thus, we have obtained a lower bound for the cost E(x u1 )2 . It can be veried that this lower bound can be achieved by linear decisions 0 and 1 (omitted here). Hence, a hard non-convex problem could be handled by using simple tools from information theory. One of the goals of this project is to use ideas such as entropy and mutual information to obtain lower bounds of the optimal costs and to construct controllers in a systematic way to get as close as possible to these bounds. Using this insight, the applicant has, in collaboration with colleagues from the communication theory lab at KTH [10], developed a generic method for ecient algorithms to tackle problems of the Witsenhausen Counterexample-character. The minimal cost obtained using our methods is 0.16692462, which is the lowest thus far!

References
[1] B. Bamieh, F. Paganini and M. Dahleh. Distributed Control of Spatially Invariant Systems. IEEE Trans. Auto. Control, 47(7), July 2002. [2] B. Bamieh and P. Voulgaris. Optimal distributed control with distributed delayed measurements. In Proc. IFAC World Congress, 2002. [3] C. Berrou, A. Glavieux, and P. Thitimajshima. Near Shannon limit error-correcting coding: Turbo codes. Proceedings of the IEEE International Conference on Communications, Geneva, Switzerland, May 1993. [4] R. G. Gallager. Low-Density Parity-Check Codes. Cambridge, MA:MIT Press, 1963.
For random variables x and y, H(x) denotes the entropy of x, H(x|y) the entropy of x given the information about y, and I(x, y) = H(x) H(x|y) denotes the mutual information between x and y.
2

Ather Gattami

Research Description

[5] A. Gattami and Richard Murray. A Frequency Domain Condition for Stability of Interconnected MIMO Systems. In Proc. of the American Control Conference, Boston, USA, June 2004. [6] A. Gattami. Optimal Decisions with Limited Information. PhD Thesis ISRN LUTFD2/TFRT--1079-SE. Department of Automatic Control, Lund University, Lund, Sweden, June 2007. [7] A. Gattami, S. Mitter. Distributed Estimation and Control under Partially Nested Information Pattern. Submitted to IEEE Trans. Automatic Control. [8] Yu-Chi Ho and Kai-Ching Chu. Team Decision Theory and Information Structures in Optimal Control Problems - Part I. IEEE Transactions on Automatic Control, 17(1):15-22, Feb. 1972. [9] U. Jnsson, C.-Y. Kao, and H. Fujioka. A Popov Criterion for Networked Systems. In Proceedings o of MTNS 2006, Kyoto, Japan. [10] J. Karlsson, A. Gattami, T. Oechtering, M. Skoglund. Source-Channel Coding Approach to Witsenhausens Counterexample. Submitted to the Conference on Decision and Control 2010. [11] C. Langbort, R. S. Chandra, and R. DAndrea. Distributed Control Design for Systems Interonnected over An Arbitrary Graph. IEEE Transactions on Automatic Control, 49(9):1502-1519, Sep. 2004. [12] J. Marshak. Elements for A Theory of Teams. Management Sci., 1:127-137, 1955. [13] P. Parrilo. Semidenite programming relaxations for semialgebraic problems. Mathematical Programming, 96:2, pp. 293-320. [14] R. Radner. Team Decision Problems. Ann. Math. Statist., 33(3):857-881, 1962. [15] A. Rantzer. Linear Quadratic Team Theory Revisited. In Proc. of ACC 2006, Minneapolis, Minnesota, USA. [16] M. Rotkowitz, R. Cogill, S. Lall. A Simple Condition for the Convexity of Optimal Control over Networks with Delays. CDC05, Seville, Spain. [17] C. E. Shannon. A mathematical theory of communication. Bell System Tech. J., vol. 27, pp. 379-423 and 623-656, Jul. and Oct., 1948. [18] S. Tatikonda, S. Mitter. Control under Communication Constraints. IEEE Trans. Auto. Control, 49(7):1056 - 1068, July 2004. [19] S. Tatikonda, A. Sahai, S. Mitter. Stochastic Linear Control Over a Communication Channel. IEEE Trans. on Auto. Control, Volume: 49 , Issue: 9 , Sept. 2004, Pages:1549 1561. [20] M. Wainwright, M. I. Jordan. Graphical Models, Exponential Families, and Variational Inference. September 2003. [21] H. S. Witsenhausen. A Counterexample in Stochastic Optimum Control. SIAM Journal on Control, 6(1):138-147, 1968.

You might also like