experiments for dynamic resource allocation, …passino/paperstopost/raschedcontrol-cs… ·...

F E A T U R EF E A T U R E

Experiments for DynamicResource Allocation,

Scheduling, and Control

dvances in infor-mation technol-

ogy such asobject-orient-

ed program-ming, real-

time operating systems, andreliable computer communi-cation networks enable thedeployment of effective con-trol and automation sys-tems [1]. Industry exploitsinformation technology indistributed control system(DCS) products, whosecomponents include pro-portional-integral-derivative (PID) and programmable logiccontrollers. At the same time, DCSs can use distributed net-worked decision-making systems for factory-wide controlsolutions. There is anongoing technological rev-olution in the applicationof networked embeddedcomputing to control. Advancements in this area are drivenby low-cost and high-performance microcontrollers thathave onboard data acquisition as well as the controller areanetwork (CAN) or wireless Ethernet technologies that inter-

connect these microcon-trollers. There are softwaretools for developing andimplementing such embed-ded distributed networkedcontrol systems. The code-sign of software and feed-back controls for both DCSand networked embeddedcontrols is a topic of currentinvestigation for control andsoftware engineers [1]. Foradditional information onDCS products and net-worked embedded controlssee “Additional Resources.”

While industry continues to exploit information technol-ogy for distributed feedback control, universities face thechallenge of revising their curricula to suit the educational

needs of future engineersin this area. We can iden-tify two research thruststhat support relevant

educational experiences. First, cooperative roboticsinvolves groups of ground, underwater, flying, or spacerobots that perform tasks such as cooperative search,cooperative pushing of an object, or coordinated group

New challenges from information technology-enabled feedback control

© PHOTODISC AND ARTVILLE, LLC.

By Nicanor Quijano, Alvaro E. Gil, and Kevin M. Passino

February 2005 630272-1708/05/$20.00©2005IEEE

IEEE Control Systems Magazine

A

February 200564 IEEE Control Systems Magazine

motion. In these applications, there is typically a micro-controller on each vehicle supported by a wireless ether-net for vehicle-to-vehicle communication. Courses thatprovide educational experiences in this area are describedin [2] and [3]. A second thrust is the research program onwireless sensor networks based on the University of Cali-fornia at Berkeley motes, which are electronic modulesthat have processor, sensor, and communication capabili-ties [4]. Networks of motes can be used for distributed net-worked control system problems, such as the tracking ofan evader moving through a field of sensors.

This article introduces a third approach to distributednetworked feedback control based on an educational labo-ratory. This laboratory emphasizes low-cost experimentsthat retain key elements of realism, problems that are dif-ferent from those studied in the cooperative robotics andsensor network areas, and new challenges from informa-tion technology-enabled feedback control [1], [5], such asdynamic resource allocation, feedback scheduling of tasks,decision making over networks, and nontraditional controlobjectives. We have developed the following experimentsto study these challenges.

● Balls-in-tubes experiment: Four balls are levitated toa common maximal height by dynamically allocat-ing air flow to four tubes that hold the balls. In onescenario, only one air pulse can be applied at atime, and the control system juggles the balls. Thejuggling is challenging, owing to significant ball-height-sensor noise, air turbulence, ball-to-ball cou-pling through a common air source, actuator

bandwidth limits, and the need for distributed deci-sion making over a network.

● Electromechanical arcade: Two agents must coopera-tively fire laser guns at photocell targets to maximizethe team’s point gain within a finite time period.Feedback scheduling is needed due to the unpre-dictable appearance and disappearance of targets.Distributed scheduling over a network makes coordi-nation challenging.

● Planar temperature grid: This experiment has a planargrid of 16 zones, each of which has a temperaturesensor, heater, and controller interface. The goal isto allocate a limited amount of electrical current tothe heaters to maximally raise the temperatures ofall of the zones to a common value. The dynamicallocation is complicated by interzone coupling,ambient temperature influences, and wind currents.Moreover, the presence of a communication networkbetween zones, with possible topological constraintson information flow, necessitates the use of distrib-uted decision making.

● Building temperature control: This experiment con-sists of a scale-model two-floor building with roomsand halls. There are heaters and sensors in eachroom, adjustable doors and windows, and electrical-ly controllable wind flows. The building has many ofthe same characteristics and challenges as the pla-nar temperature grid problem, but different inter-zone and ambient temperature effects.

In this article, we identify the role of these experimentsin a university curriculum. The choice of laboratory soft-ware and hardware is discussed, followed by an introduc-tion to the experiments. For each experiment, we describethe apparatus, the challenges that must be confronted tomeet performance objectives, sample feedback controlsolutions, and alternative control methods that might beuseful. In the concluding remarks, we explain how the mod-ular nature of each apparatus provides opportunities toexpand the experiments to confront additional challenges.

Laboratory Role in a University CurriculumThe Department of Electrical and Computer Engineering atOhio State University (OSU) has a modern undergraduatecontrol laboratory that uses dSPACE hardware and soft-ware [6] and Quanser experiments [7]. Within the OSUquarter system, we have two ten-week, dual-level gradu-ate/undergraduate laboratories designated EE 757 and EE758. Neither EE 757 nor EE 758 is a prerequisite for theother. The majority of students taking EE 757 and EE 758are graduate students who have had at least one year ofgraduate course work in the control area. Students comeprimarily from the electrical, mechanical, industrial, andaerospace engineering disciplines.

Additional ResourcesDistributed Control System (DCS) ProductsThere are a wide range of companies that sell DCS prod-ucts. Major vendors include: Honeywell, Yokogawa, Emer-son, Invensys Foxboro, Rockwell Automation, and ABB.See the Web sites of these companies or any issue of theControl Magazine: http://www.controlmagazine.com/.

Networked Embedded SystemsThere are many companies that sell products for embeddednetworked control systems. See, for instance, the Web sitesof Texas Instruments, Motorola, or Microchip Technology.For a product directed at industrial solutions, see the Hon-eywell “Smart Distributed Systems” product at: http://con-tent.honeywell.com/sensing/prodinfo/sds/.

OSU Distributed Dynamical Systems LaboratoryFor more information on the experiments described in thisarticle, including videos of their operation, seehttp://www.ece.osu.edu/ ∼passino/distdynamicsyslab.html.

February 2005 65IEEE Control Systems Magazine

EE 757 focuses on cooperative robotics problems suchas platooning and cooperative search [2]. Small ground-based robots, connected by means of a wireless ethernet,are used as the experimental testbed. EE 758 is a servicecourse that provides students with laboratory experiencesin introductory linear, nonlinear, and robust control, alongwith more specialized topics [8]. The first half of EE 758 isdedicated to learning dSPACE, data acquisition, modelingand system identification, PID control with antiwindupschemes, the linear-quadratic regulator with an observer,and nonlinear control using feedback linearization. Theexperiments used in the first half of the course areQuanser products [7], [8].

The second half of EE 758 is dedicated to in-depth studyon a topic of interest to the student. A faculty member candevelop suitable experimental challenges for students whofocus on system identification, sliding mode control, adap-tive control, intelligent control, game-theoretic approach-es, advanced robust control, or advanced nonlinearcontrol. Additional Quanser experiments are available forthe second half of EE 758, including tanks, a two-degree-of-freedom helicopter, and an inverted cube [7]. Moreover,we can provide students with projects that use the Nation-al Institute of Standards and Technology (NIST) Real-TimeControl Systems (RCS) software library [9] or NationalInstruments hardware and LabVIEW software [10].

The default project sequence in the second half of EE758 is distributed networked feedback control. We typicallygive the students one of the experiments described in thisarticle and require them to design, implement, and testresource allocators, schedulers, and controllers. We oftenrequire the students to design strategies that can overcomeinstructor-induced disturbances, such as a puff of air on themultizone temperature control experiment. While such dis-turbances are difficult to quantify and repeat, they provideclear challenges and tangible phenomena for evaluating thestudent’s understanding of distributed networked feedbackcontrol. Students are expected to provide a demonstrationand written report for all experiments.

Laboratory Software and HardwareSuccessful development of an educational laboratory fordistributed networked feedback control depends on thechoice of software, processor, and data acquisition hard-ware. One approach is to purchase DCS software and hard-ware products that are used in industry, which has theadvantage of familiarizing students with state-of-the-arttools. Unfortunately, many industrial DCS products are tooexpensive for universities to purchase and maintain. More-over, some DCSs lack flexibility and have a market nichethat can lead to an over-specialization of student skills,resulting in a distraction from foundational principles.

An alternative is to use a networked embedded controlapproach, with multiple microcontrollers interconnected

by means of a CAN bus. For this approach, competing soft-ware products must be evaluated. The networked embed-ded control approach has the advantage of being relativelyeasy and inexpensive to expand to a large number of deci-sion-making nodes in some applications [11]. However, forsome experiments, this approach can have the disadvan-tage of not facilitating comparisons to centralizedapproaches, which is often useful in research and educa-tion. Similar comments apply to the motes in [4].

Another alternative is to use software and hardwaresolutions that were developed for implementing a singlecontroller but can support distributed control. For exam-ple, dSPACE [6] provides controller boards and softwarebased on Mathworks’ MATLAB and Simulink [12]. In thedSPACE products, there are a number of ways to communi-cate between controllers hosted by different computers,including a serial link or the Internet. The dSPACEapproach has the advantage of building on students’ MAT-LAB skills gained from courses on signal processing andcontrol design. Moreover, each laboratory station at OSUhas a computer with resident dSPACE hardware and soft-ware. When laboratory courses are not being offered, wecan use the hardware and software for experimentalresearch on distributed networked feedback control. Analternative to dSPACE is National Instrument’s hardwareand LabVIEW software [10].

An attractive but less commercially popular alternativeis to use the NIST RCS software library [9], [13], [14]. TheNIST RCS software was developed to provide the followingfeatures: platform-independent software for use on differ-ent computers with different operating systems; softwaresupport for a range of communication technologies andmessage formats; design support for hierarchical and dis-tributed control modules; and a graphical user interfacefor online monitoring and debugging. The NIST RCS soft-ware is under continual development and is used in arange of applications [9], [13]. The primary disadvantageof the NIST RCS software is the lack of an interface to MAT-LAB and, in some cases, the need to develop an interfaceto data acquisition hardware. A significant advantage isthat a mature design methodology for distributed net-worked feedback control is available [13].

OSU has been funded for a number of years by NIST totransfer RCS software and methodology to a universityeducational setting [13], [14]. Recently, NIST funded anOSU evaluation on the potential of transferring the RCSdesign methodology to dSPACE and National Instrumentsproducts. Based on these studies, our solution to the soft-ware/hardware choice problem is to use dSPACE productscoupled with the NIST RCS design methodology. Thisapproach allows exploitation of MATLAB in the lab, sup-ports pedagogy for design methodology, allows us to inte-grate the hardware/software resources from each labstation, and provides support for comparisons between


distributed and centralized feedback control solutions.Since the full details of RCS are covered elsewhere [9],[13], [14], we only briefly describe its adaptation for use inour laboratory with dSPACE software. Adaptation of RCSmethodology for LabVIEW is conceptually similar.

RCS software has design and diagnostic tools, whichcan be used to develop distributed control algorithms andmonitor and change variables in real time. The functionali-ty of these tools can be implemented in dSPACE productsthrough Simulink and the dSPACE graphical user interface.To develop the control system block diagrams in Simulinkfor the RCS design methodology, we use preprocess, deci-sion process, and postprocess submodules in a controlmodule. Then, several control modules can be connectedin a hierarchical and distributed fashion. In the pre-process submodule, we acquire data, perform conver-sions, and store information in global variables for use inthe decision process submodule. Next, we develop severalsubsystems that are responsible for making control deci-sions or other tasks such as updating variables used in the

postprocess submodule, which transmits data to othercomputers and updates the digital outputs.

The Balls-in-Tubes ExperimentThe balls-in-tubes experiment is designed to be an inex-pensive testbed for dynamic resource allocation strategiesthat exploit information from the plant in feedback.

Experimental Apparatus and ChallengesFigure 1 shows the balls-in-tubes experiment. There arefour modules, each of which has a tube with a ball inside, afan at the bottom to lift the ball, and a sensor at the top tosense the ball’s height. For each tube, there is a box thatthe fan pressurizes. The box can be viewed as a stiff bal-loon that is blown up by the fan and whose outlet holeblows air into the tube. The tubes are connected at the faninlets by means of an input manifold that has an inlet atthe bottom, as indicated in Figure 1. Each tube has an out-put manifold at the top with an outlet as shown.

The manifolds are a key component of the experiment,as they necessitate the sharing ofair at the input or restrict air flowat the output; both actions causesignificant coupling among the fourtubes. Characteristics of the cou-pling can be adjusted by designingdifferent opening sizes for the inletand outlet. Alternatively, solidobjects can be placed inside theinput or output manifolds toobstruct airflow. The air flow char-acteristics are complicated due toturbulence in the manifolds, pres-surized box, and tubes. For a rangeof input manifold inlet sizes, a fansucceeds at lifting the ball only atthe expense of other balls drop-ping. This feature leads to the needfor resource allocation, where theresource is the air that elevates theballs. The input manifold effective-ly implements a multivariable satu-ration constraint. Under someconditions, a fixed amount of aircoming from the input manifoldcan be allocated. When significantamounts of air are allocated tosome tubes, the fans for the othertubes become limited in what theycan allocate.

To measure the ball heights, weuse Devantech SRF04 ultrasonic sen-sors. The raw data obtained fromthe sensors are quite noisy, with

Figure 1. Balls-in-tubes experiment. The tubes are numbered from left to right. Eachtube has a dc fan, a ball, and an ultrasonic sensor. There is an input manifold inlet atthe bottom and an output manifold outlet at the top. These manifolds necessitate thesharing of air at the input, or restrict air flow at the output.

Tube 1

Ball 1

Close-Up View of Output Manifold

UltrasonicSensor 3

Output ManifoldOutlet

PressurizedBox 1

Close-Up View of Input Manifold

dc Fan 4

Input Manifold

Inlet

1710

mm

465 mm

470 mm


typical measurement spikes corresponding to errors greaterthan 10 cm. Using digital filters to smooth the sensor out-puts, we achieved a resolution of ±1 cm. The actuators areDynatron DF1209BB dc fans, commonly found inside person-al computers. We use a pulse-width-modulation (PWM) sig-nal as an input to the fan. The sampling period is 100 µs, andthe period for the PWM is 0.01 s. There is one digital inputand one digital output for each sensor and one digital outputfor the PWM input to each fan, for a total of 12 digital input-output lines that connect to a DS1104 dSPACE card.

The total cost of the plant is less than US$200 for fourtubes, including the sensors and actuators. The supportingelectronics for the actuator and sensor interfaces are rela-tively inexpensive. Tube modules designed to be easilyseparated can be used at a separate laboratory station tostudy the control of ball height in a single tube. A projecton ball height control in a single tube might be a usefulchallenge, perhaps even in an undergraduate laboratory.

In addition to balancing a single ball in a tube, thereare a number of control objectives and challenges thatcan be studied for this experiment:

● balancing the balls inside the tubes, trying to allo-cate air flow to keep all the balls at fixed positions ormaximally elevated at a uniform height

● balancing and reallocation in the presence of plantchanges due to manifold inlet size changes or flowobstructions in a manifold

● effects of distributed networked decision making inthe presence of an imperfect communication network.

Resource Allocation Strategies and ResultsWe describe two resource allocation strategies and ana-lyze their performance in the presence of an inlet sizechange or an obstruction placed inside the input manifold.

Resource Allocation Strategy: JugglerThe goal of the juggler resource allocation strategy is tomaintain the balls in the vicinity of a specified commonheight. Although the peak-to-peak oscillation of each ballheight can be relatively large, we require the time averagesof the heights of each of the four balls to be the same. Wecomplicate this goal by allowing only one fan at a time tohave a sufficient portion of the PWM duty cycle to raisethe corresponding ball. This constraint leads to a jugglingresource allocation strategy that is designed to minimizethe differences between the average ball heights despiteair turbulence, intertube coupling through the manifolds,fan bandwidth constraints, and significant sensor noise.The dynamics of allocation rely critically on feedbackinformation on ball heights and the distributed decision-making that implements the allocation strategy.

The following characteristics of the experiment are rele-vant to the design of the juggler strategy.

● The time it takes for a ball to rise a fixed distance ina tube is generally greater than the time it takes theball to drop the same distance.

● Since the inlet to the input manifold is closest to thefan inlets to tubes 2 and 3, air flow for these twotubes is less restricted than for tubes 1 and 4. Hence,for tubes 2 and 3, the amount of time it takes to raisethe balls or have them drop is less than the corre-sponding times for tubes 1 and 4.

● Apparently due to module construction differences,the time it takes to lift the ball in tube 4 is generallygreater than the corresponding time for tube 1.

● The minimal percentage of the duty cycle of thePWM signals that can lift all of the balls simultane-ously is about 60%. This input is just strong enoughto pressurize the boxes so that the air pressure canlift the balls.

For fan i at time t, let Diu(t) and Di

d(t) be the duty cyclesthat cause the ith ball to go up or down, respectively. Wechose these values using the previous insights and sometrial-and-error tuning. The juggler strategy consists of thefollowing steps:

1) Find a ball to focus on: Let i ∗ be the lowest ball attime t and hi∗(t) be its height.

2) Push the low ball up: For all t′ ∈ [t, t + T], for somefixed time period T , assign Di∗

u (t′) to fan i ∗ and, forj �= i∗ , assign Dj

d(t′) but modify the input based onthe following rules• Truncation rule 1: If hi∗(t′) > 0.6 m at time

t′ ∈ [t, t + T], then the algorithm aborts the cur-rent allocation and returns to step 1, even if thetime duration T has not elapsed. This rule avoidspushing ball i ∗ too high.

• Truncation rule 2: If hi(t′) > 0.6 m for all i at timet′ ∈ [t, t + T], then the algorithm aborts the cur-rent allocation, assigns Di

d(t) to each fan for 0.6 s, and then goes to step 1. This rule helpsavoid having all of the balls get stuck at the top ofthe tubes, which can happen in some cases if theinlet to the input manifold is not restricted in size.

In summary, the juggling strategy involves dynamicallyapplying different size and duration air pulses to the tubescorresponding to the minimum ball height. We tuned thevalue of the time duration T until we found that T = 3 swas acceptable.

Figure 2 illustrates the closed-loop performance of thejuggler strategy. The oscillations in ball heights are a directconsequence of the constraint that only one fan at a timecan be given an input that can lift the corresponding ball.The average heights of the balls are reasonably close. Wecould tune the strategy to obtain a closer correspondencebetween the averages of the ball heights, however, such tun-ing can be tedious and futile for the following reasons. First,the module differences that arise from sensors, actuators,


air leaks, and physical construction are dynamically cou-pled through turbulent air flows. Tuning the juggler strategyaffects this unpredictable coupling, and the small differ-ences in average ball heights emerge. Second, we have seentemperature changes in the room affect the performance.

A typical feedback control goal, such as zero steadystate tracking errors between the ball heights and a com-manded height, is simply not achievable for the given con-straints. Moreover, the achievable common average ballheight is not known a priori; hence, the problem is not one

of trying to regulate the average ball heights to a command-ed value. The closed-loop performance objective for theexperiment is nontraditional but nonetheless quite natural,as most people have seen a human juggler perform andunderstand that juggling is a feedback control problem.

Next, we conducted an experiment using the jugglerstrategy, but with an unmodeled plant change. To changethe plant, we placed an object inside the input manifoldto partially obstruct the air flow. The obstruction espe-cially affected tube 4 since the object was placed beneath

its fan. The plant change necessi-tated reallocation of the air flowrelative to the nominal case dis-cussed previously. To illustratethe required reallocation, we ranthe experiment with and withoutthe obstruction and computed theaverage ball heights. We also com-puted the percentage of the 200 sexperiment run time that each ballwas pushed up. In other words,we computed the fraction of timethat Di∗

u (t) was applied to eachball; see Table 1 for the results.Due to the characteristics of theexperiment, the allocator attendsto tube 4 the most. The averagevalues for the ball heights arecomparable to the ones for thenominal case but generally lowerdue to the flow obstruction. Thekey feature is how the amount oftime allocated to tube 4 increasesto compensate for the air flowobstruction. Basically, the strate-gy takes some of the air that wasallocated to tubes 2 and 3 in thenominal case and reallocates it tomaintain the relatively equal aver-age heights. This phenomenon isthe essence of dynamic feedbackresource allocation.

Resource Allocation Strategy: Dynamic ProportioningFor the dynamic proportioningresource allocation strategy, wefirst design inner-loop ball heightcontrol systems for each tube. Weuse a simple proportional controllerthat has a static nonlinear gain onthe error ei(t) = hd(t) − hi(t), wherehi(t) is the ball height for the ithtube and hd(t) is the desired height.

Figure 2. Heights of the balls in each tube for the juggler strategy. The oscillations inthe ball heights are expected since only one fan at a time is given an input sufficientto raise the corresponding ball. The indicated average heights are reasonably close.The standard deviations of the heights are 0.14, 0.17, 0.20, and 0.15 m for tubes 1through 4, respectively.

0

0.2

0.4

0.6

0.8

1

0

0.2

0.4

0.6

0.8

1

0 50 100 150 2000

0.2

0.4

0.6

0.8

1

0

0.2

0.4

0.6

0.8

1

Hei

ght,

h 3(t

), m

Hei

ght,

h 4(t

), m

Hei

ght,

h 1(t

), m

Hei

ght,

h 2(t

), m

0 50 100 150 200

t, s

Mean

MeanMean

Mean

Table 1. Allocation results for the nominal and modified plants. Note that theobstruction to flow for tube 4 demands a reallocation that results in moreattention given to tube 4 to render the values of the average ball heights closeto each other. A side effect is that the average ball heights are generally lowerthan in the nominal case.

Nominal Plant Plant with Obstruction

Percent of time Percent of time each tube Height each tube is Height avg.is attended avg. (m) attended (m)

Tube 1 23.22 0.62 24.41 0.59

Tube 2 10.95 0.72 6.98 0.64

Tube 3 22.69 0.70 13.61 0.71

Tube 4 43.13 0.61 54.99 0.51


Although a PID controller might perform better, such designimprovements for the inner-loop ball height control systemare not our focus. Our proportional controller changes theduty cycle of the PWM used to drive the fans, and its gaindepends on the sign of the error. The controller will haveone slope if the error ei(t) is positive, and another slope ifthe error is negative. The two slope values are chosen sothat the balls do not drop so fast that they are too difficultto lift at a later time due to inertia.

For the outer-loop resource allocation strategy, wechange hd(t) every 10 s by letting

hd(t) = max{hi(t) : i = 1, 2, 3, 4}.

This strategy directs the fan corresponding to the lowestball to provide sufficient air to lift the ball to the sameheight as the highest ball. A time of 10 s is sufficient toachieve the modified ordering. The dynamic iterative appli-cation of this simple allocation rule results in elevating theballs to a maximum uniform height. It is essentially impossi-ble for the balls to be at exactly the same height in imple-mentation, owing to noise, air turbulence, and ballovershoot. One ball will be higher than the others at virtual-ly all times. The outer-loop resource allocation strategycommands the lower balls to achieve the position of thehighest ball. Over time, the strategy persistently drives theballs to successively higher uniform positions. Whatemerges is a proportioning of airflow with relative amounts of airallocated to overcoming flow differ-ences between the tubes, as well asdisturbances and plant changes.Since the resource allocation strate-gy proportionally allocates thePWM input, this strategy is differentfrom the juggler strategy. The jug-gler allocates air pulses over time,whereas the dynamic proportioningstrategy continuously allocates airboth spatially across the tubes andtemporally due to the changes inthe commanded value of hd(t).

To test this allocation strategy,we limit the amount of air enteringat the inlet of the input manifold;otherwise, all of the balls will hitthe top. To restrict flow at the inlet,a piece of paper with a hole cut inthe middle is used to cover the airinlet. In the following, we describehow holes of different sizes in thepaper affect dynamic reallocation.

As shown in Figure 3, all the balls reach the desiredheight hd(t), except the first one. The balls converge to acommon height of around 0.7 m and after two minutes,their heights are within approximately 0.2 m of each other.However, even though the ball in tube 1 is in the 0.2 mrange, ball 1 does not reach the maximum desired heighthd(t). This shortcoming is due to several factors. First, airenters in the middle of the input manifold, which is closeto tubes 2 and 3. The inner-loop ball height controller forthe height of ball 1 continually tries to obtain more air butis at a disadvantage due to the flow restriction. Althoughwe attempted to overcome the differences between thetubes by tuning the inner-loop proportional controllers, wedid not dedicate much time to tuning since the inner-loopcontrols were not our main focus. Adding an integral termto the inner-loop controllers might have reduced the errorseen in Figure 3. However, since we did not try a PI con-troller and since there is strong coupling between theinner- and outer-loop controls, we cannot be certain thatsuch improvements could be obtained. One dominant fac-tor in this coupling is air flow. We cannot measure the pat-terns of air flow inside the manifold, and the flow patternsmake the dynamics inside the manifold change unpre-dictable due to turbulence. Some effects of air turbulenceare easy to see while the experiment is in operation. Forinstance, even with constant PWM signals to the fans, theballs bounce and spin during operation. We find that theair turbulence creates significant coupling between theinner- and outer-loop controls and hence limits the

Figure 3. Dynamic proportioning strategy results. The figure shows the ball height ineach tube and the desired height hd(t), represented by SP. The desired height changesevery 10 s to the height of the highest ball.

0.2

0.4

0.6

0.8

1

Hei

ght,

h 1(t

), m

0.2

0.4

0.6

0.8

1

Hei

ght,

h 3(t

), m

0.2

0.4

0.6

0.8

1

Hei

ght,

h 4(t

), m

0.2

0.4

0.6

0.8

1

Hei

ght,

h 2(t

), m

0 50 100 150 200 0 50 100 150 200t, s

SP

SP SP

SP


achievable performance. Such features have been studiedin a variety of applications and are a key challenge indesigning distributed and hierarchical controllers forlarge-scale systems [13], [15], [16].

Figure 4 illustrates the results of dynamic reallocationwhen the inlet area is reduced at t = 20 s by suddenly cov-ering the inlet with a piece of paper with a small hole cut init. Since there is not as much air to proportion after theplant change, the balls fall. However, the balls fall some-what uniformly owing to the dynamic reallocation. For thiscase, we had to include a constraint that indicates thatwhen all of the balls are below 0.2 m, the desired height isset at a fixed value of 0.4 m. This rule is used in some runsof the experiment to keep the balls from getting stuck atthe bottom.

The necessity for such heuristic rules highlights theneed to understand the theoretical underpinnings of the

dynamic resource allocation strategy. We need to modelthe plant and analyze the choice of hd(t) and the effects ofdynamics and noise on ultimate achievable height andheight uniformity. Resource allocation was originally con-

ceived as a static optimization prob-lem [17]. Extending resourceallocation concepts to feedback con-trol systems requires innovations inresource allocation theory to copewith dynamics and disturbances.

Electromechanical ArcadeThe electromechanical arcadeexperiment is designed to be aninexpensive testbed for networked

scheduling strategies where information must be sharedto achieve success.

Experimental Apparatus and ChallengesThis experiment is composed of two main components:guns mounted on the shafts of motors and targets (Figure5). Each of the two guns has a laser and a photodetector.To fire a gun, the gun’s laser is activated by a computer.There are eight targets, which are black boxes arranged inan arc. Each target has a single photodetector and twolasers mounted on top, with a laser pointed at each gun.We represent the appearance and disappearance of targetsby whether the target’s two lasers are both on or both off.

When a target appears, we consid-er it to have “popped up.” A guncan sense the target’s appearanceby detecting its laser, but only ifthe gun is pointed directly at thattarget by its motor. A gun cannotdetect the appearance of morethan one target at a time. If a gunfires at a target that is currentlypopped up and the gun’s laser suc-cessfully triggers the photodetec-tor on the target, then the gunreceives a point for hitting the tar-get. However, the targets also dis-appear and if the target is firedupon at that time, the gun receivesno points. The appearance fre-quencies of the eight targets areindependent of each other. The tar-get lasers are driven by simple tim-ing circuits that can be adjusted byhand to provide different frequen-cies. We use Radio Shack #277-1101lasers and Radio Shack #276-1657photodetectors for both the tar-gets and guns.

Figure 4. Dynamic reallocation results. The figure shows the ball height in each tubeand the desired height hd(t), represented by SP. The desired height changes every 10 sto the height of the highest ball. Here, the inlet area is reduced at t = 20 s to demon-strate dynamic reallocation.

0

0.2

0.4

0.6

0.8

1

1.2

Hei

ght,

h 1(t

), m

0

0.2

0.4

0.6

0.8

1

1.2

Hei

ght,

h 2(t

), m

0

0.2

0.4

0.6

0.8

1

1.2

Hei

ght,

h 3(t

), m

0

0.2

0.4

0.6

0.8

1

1.2

Hei

ght,

h 3(t

), m

0 50 100 150 200 0 50 100 150 200t, s

SP

SP SP

SP

Experiment modularity facilitatesthe construction of new configurationsfor additional research andeducational objectives.

The two guns are mounted on the shafts of ShinanoKenshi #LA052-040E4N02 motors driven by AdvancedMotion Control #BE12A6 amplifiers. Each motor has aquadrature encoder to measure the angular position ofthe motor. We limit the angular velocity of each motorshaft to make the problem more challenging and to makeit possible to view the firing of guns at targets. We use aPID controller to point each gun to the target that needsto be shot, and these PID control system loops are tunedso their performance does not affect our schedulingstrategies. The data acquisition system requirements areten digital inputs to the computer from the target and gunphotodetectors, two digital outputs to fire the gun’slasers, two motor shaft encoder interfaces, and two ana-log outputs for the PID controller inputs to the motoramplifiers. We use the DS1104 dSPACE card.

The cost of the experiment is relatively low, with thegreatest expenses occurring from the motors and ampli-fiers. However, the specifications for the motors are notoverly demanding since high torques are not needed. Rel-atively low shaft rotation speeds are desirable to allowviewing of the arcade. Required gun pointing accuracy isdriven by the laser’s light dispersal pattern and the size ofthe photodetector. As long as the targets are kept close tothe guns, highly accurate pointing is not needed. In fact,the target appearance timing circuits can be eliminated ifadditional digital outputs are available as they are on the

DS1104. Computer control of the targets provides a flexi-ble way to command the pattern of appearances.

We assume that the guns have no information about therates of appearance of the targets, although strategies couldbe invented for estimating appearance sequences. The gunsknow the positions of all of the targets, and the guns cancommunicate their decisions to process or pursue targets.The challenges for this experiment are as follows:

● to schedule a sequence of firings in real time to maxi-mize the number of points received by the team of twoguns. Since target detection and shooting requiresmovement of the guns, a good schedule typically min-imizes the motion of the guns and, at the same time,maximizes point gain. Feedback is required to over-come uncertainty about when targets appear and todevelop target appearance time estimation strategies.

● to cooperatively schedule the shooting of the guns in thepresence of imperfect communication between the twoguns. While the communication network can be theInternet, and a computer can be dedicated to eachgun, networked schedulers can also be simulatedwithin one computer. Communication imperfectionssuch as random but bounded delays, bandwidth con-straints, or message misordering can be considered.

Finally, this distributed feedback scheduling problemcan be thought of as a resource allocation problem, analo-gous to the one faced by the juggler. However, the dynam-


Figure 5. Electromechanical arcade experiment. The experiment has two guns and eight targets. The guns detect the targetsby sensing lasers mounted on the targets. Each gun receives points by firing its laser at targets when the targets appear. Eachgun is mounted on a motor that is under computer control so that the gun can be pointed at different targets.

Gun 1

Gun 2

Guns Firing Targets

Target 1

Motion ControlAmplifier for

Motor 1

Motor 1

Motion ControlAmplifier for

Motor 2Motor 2

Laser 1of Target 4

Laser 2of Target 4

Photodetector of Target 7

Laserof Gun 2

Photodetectorof Gun 1

ics, uncertainty sources, and performance objectives arequite different.

Scheduling Strategies and ResultsThe arcade experiment involves an uncertain environmentdue to unpredictable target appearance times. Imperfectcommunication makes it difficult for the two guns to coor-dinate their actions. Because of uncertainty, it is not gener-ally possible to accurately predict far into the future, andhence optimization approaches are not effective for devel-oping long sequences of target firings either offline oronline. Hence, we develop a different approach here.

The targets are numbered, the set of targets is denotedby P = {1, 2, . . . ., 8}, and the set of guns is denoted byQ = {1, 2}. Let pi, i ∈ P, denote the priority of processingtarget i, where processing means moving to and shootingat the target when it appears. The priority pi > 0 is propor-tional to the importance of a target. Let Ti(t) = piti, wherei ∈ P and t ≥ 0, denote what we call the “prioritized time”elapsed since target i was last shot, where ti denotes theamount of time elapsed since target i was last shot. Sup-pose that, initially, Ti(0) = 0 for all i ∈ P so that we act asthough we had simultaneously shot all of the targets,which is clearly physically impossible. However, this ini-tialization is useful since the scheduling strategies makedecisions about which target to process based on the sizesof the prioritized times Ti(t) for all i ∈ P.

If no gun is used, then, for all i ∈ P , Ti(t) → ∞ ast → ∞, since no targets are processed. The goal of ascheduling strategy is to avoid Ti(t) → ∞ for all i ∈ P.Moreover, a successful strategy tries to keep the values ofTi(t) as small as possible to reflect that the guns haverecently shot each target.

There are three types of delays in the experiment thathave a significant impact on the choice of scheduling strate-gy. First, there is the delay between the appearances of eachtarget. Second, there is travel time to move the gun to pointat a target. Third, there is a random but bounded delayincurred for each communication between the guns. Weimplement an asynchronous scheduler over a simulatedcommunication network to pick i∗j (t), the target that gun j iscommanded to process at time t. To define the scheduler, letU(t) ⊂ P be the set of unattended targets and letUa

j (t) = {i∗j (t)} ∪ U(t) be the set of targets that gun j can con-sider after processing (shooting at) target i∗j (t). We pass theset U(t) around the network, let each gun choose a targetfrom Ua

j (t), and include in U(t) the target that gun j just fin-ished processing. A distributed mutual exclusion algorithmis used, and a communication delay is incurred each timeU(t) is passed. While a gun is waiting to receive U(t), the guncan keep shooting at i∗j (t) but receives only one point perappearance of the target. Hence, we consider i∗j (t) to beattended to until gun j makes the decision and switches toprocessing a different target.

Distributed Scheduling Strategy: Focus on the Target Ignored the LongestNext, we introduce a scheduling strategy from [18] and[19] that is an extension of a strategy defined in [20] andfor which stability properties have been investigated in[19] and [21]. First, let Dj denote the decision time for gunj. At each decision time, gun j chooses a target to process.Only one gun can choose a target to process at each timedue to the mutual exclusion algorithm. Suppose thatDj = 0 for all j ∈ Q so that at t = 0 one gun chooses whichtarget to process.

One candidate scheduling strategy chooses to process atarget if the target is ignored longer than the average of allthe lengths of time that the targets were ignored. Under sucha strategy, gun j chooses to process target i∗j (Dj) at time Dj if

Ti∗j (Dj) ≥ 1N − M + 1

∑

ij∈Uaj (Dj)

Tij(Dj).

Here, N = 8 targets and M = 2 guns. Gun j makes no otherdecision until it has finished shooting target i∗j andreceives U(t). A special case of this process is a strategythat picks i∗j (Dj) such that Ti∗j (Dj) is bigger than everyother Tij(Dj), i.e., a process-the-target-ignored-the-longeststrategy. If there is more than one maximizer for any gun atany time, then the strategy simply chooses one of thesetargets at random.

Figure 6 shows the results for the process-the-target-ignored-the-longest strategy. In the following, we use afixed communication delay of 10 s to facilitate comparisonto the approach. Figure 6 shows the values of Ti(t) for alli ∈ P as well as the targets selected by the two guns duringthe time the experiment is running. First, consider the ini-tial choices made by the guns. Before the experimentstarts, both guns are pointing at target 1. Once the experi-ment starts, gun 1 chooses a target to process, while gun 2must wait 10 s for the arrival of the set U(t) sent by gun 1.During this 10 s interval, gun 2 detects target 1, fires on itand receives a point, and keeps processing target 1. There-fore, T1 = 0 in this interval.

Next, consider the pattern of the trajectories in Figure 6.First, consider the effect of the priority values pi on thevariable Ti . We assigned to the targets the prioritiesp1 = 0.9, p2 = 0.85, p3 = 0.8, p4 = 0.75, p5 = 0.4, p6 = 0.4,p7 = 0.5, and p8 = 0.2. These values were chosen byobserving the illumination frequencies of each target andassigning higher priorities to the targets that appear morefrequently. With these priorities, the guns return more fre-quently to targets that appear more frequently. Second,consider the effects of the communication delay on theperformance of the strategy. For instance, notice how thevalues of Ti contained in the set U(t) are ignored duringthe communication delays. Third, notice how the guns


allocate their processing capabilities to shoot and gainmore points for as many targets as possible in a coopera-tive manner. Cooperative behavior can be seen by thesequence of choices over the time range in the latter halfof the experiment in Figure 6.

Distributed Scheduling Strategy: Focus on Closest Highest Priority TargetWe now introduce a strategy that schedules the next targetto avoid ignoring high priority targets for too long and, atthe same time, minimize travel time to process targets. Weview the electromechanical arcade experiment as analogousto a type of cooperative scheduling for autonomous militaryvehicles firing on targets. From this perspective, it is impor-tant to pay attention to minimizing the travel time betweentargets. This consideration motivated us to define the strat-egy [21] that the gun processes target i∗j (Dj) at time Dj if

Ti∗j (Dj) − δji∗j (Dj) ≥ 1N − M + 1

∑

ij∈Uaj (Dj)

[Tij(Dj) − δjij(Dj)

].

(1)

Here, N = 8, M = 2, and δjij is the amount of time requiredfor the gun to move from target j to target ij. Choice of themaximum value on the lefthand side of (1) results ina process-the-closest-high-est-priority-target policy.Ties are broken with anarbitrary choice. The leftside of (1) can be viewedas a cost function, and thescheduler selects a targetin order to minimize thetime the targets areignored and the traveltime. A bound on the ulti-mate longest time that anygun ignores target i isgiven in [21].

Figure 7 shows theresults for the process-the-closest-highest-priori-ty-target strategy. As inthe previous case, thecommunication delay isfixed and equal to 10 s,and the priority valuesare the same. The num-ber of points in this caseis greater than or equal tothe number of pointsobtained in Figure 6. Fig-ure 7 shows that the

value of the peak for each Ti is less than the peaks seen inFigure 6. By attempting to minimize travel time, weimprove our score. Data in Figure 7, specifically in thetime range t ∈ [0, 25], show that the closest targets arechosen by the guns more frequently than those for thecorresponding time range in Figure 6.

We quantify the characteristics highlighted previouslyby using performance metrics. There are many ways tomeasure performance of the scheduling strategies. Here, ateach step k, we compute the average (1/N)

∑Ni=1 Ti(k) of

the time elapsed since any target has been shot. First, wecompute the time average of this quantity, which is thetime average of the average values. In addition, we com-pute the maximum average value achieved over the entirerun of 100 s. At each time step k, we compute the maxi-mum time maxi{Ti(k)} that any target has been ignored.Second, we compute the time average of this quantity,which is the time average of the maximum values. Next, wecompute the maximum of the maximum values achievedover the entire run. Finally, we compute the number ofpoints obtained by the two guns during the experiment,which is the total number of targets shot.

Table 2 shows the results for the two scheduling strate-gies. The process-the-closest-highest-priority-targets


Figure 6. Performance of the process-the-target-ignored-the-longest strategy. At time t = 0,both guns are pointing to target 1. There is a fixed communication delay of 10 s. The values ofTi plotted versus t show the prioritized time at which target i ∈ {1, . . . , 8} was last shot. In theplot on the lower right-hand corner, the solid line corresponds to i*1 (t), and the dashed line cor-responds to i*2 (t).

0

10

20

30

0 50 100 0 50 100 0 50 100

5

0

10

t, s

T1 T2 T3

0

10

20

30

T4

0

10

20

30

T7

0

10

20

30

T8

0

10

20

30

T5

0

10

20

30

T6

0

10

20

30

0

10

20

30

i 1(t

), i 2

(t)

**

strategy performs better for all performance measures,including the total number of points obtained. Seeking tominimize travel time allows significant point improve-ment. The strategy does not, however, partition the tar-gets so that each gun is always responsible for a fewadjacent targets. What emerges in the bottom right plotin Figure 7 is a type of cooperation that balances theobjectives quantified in (1) for the process-the-closest-highest-priority-targets strategy.

A number of other meth-ods can be tested on theelectromechanical arcade.In addition to extensions ofthe strategies in [20], thereare manufacturing systemscheduling methods, bothdeterministic and stochas-tic, that might be useful forthis experiment [22], [23].Moreover, it would be inter-esting to include manualcontrol and challenge twoteenagers to beat the abovestrategies for this game.

MultizoneTemperatureControlTemperature controlexperiments are perhapsthe least expensive ex-periments that possessinteresting distributednetworked control prob-lems. Here, we brieflydescribe two multizonetemperature control ex-periments, namely, a pla-nar grid of sensors andheaters, and a buildingtemperature control prob-

lem. These experiments include traditional multivariabletracking objectives, as well as objectives found in dynamicresource allocation problems.

Experimental Apparatus and ChallengesOur first experiment has 16 individually controlled zonesarranged in a planar grid as shown in Figure 8. Each zone hasa light for a heater and an analog temperature sensor. Ambi-ent temperature variations and wind currents have a signifi-

cant impact. Since some of the heatersare close to other zones’ sensors, thereis significant inter-zone temperature cou-pling. Moreover, experiments show thatair currents can be generated by heater-induced spatial temperature gradients.

There are several practical chal-lenges in implementing this experiment.First, since each DS1104 dSPACE cardsupports only eight analog inputs, weneed two computers to acquire the datafrom the sensors. Sensed informationand actuation signals are transmitted


Table 2. Performance measures for the arcade. The process-the-closest-highest-priority-targets strategy is superior for each performance measuredue to its persistence in seeking to fire on nearby targets.

Ignored-the-Longest Closest-Highest-Priority

Average of average 7.28 5.22

Maximum of average 11.15 8.69

Average of maximum 14.46 11.56

Max of max 23.87 20.21

Number of points 25 35

Figure 7. Performance of the process-the-closest-highest-priority-targets strategy. At t = 0, bothguns point to target 1. There is a fixed communication delay of 10 s. The values of Ti plottedversus t show the prioritized time at which target i ∈ {1, . . . , 8} was last shot. In the plot in thelower right-hand corner, the solid line corresponds to i*1 (t), and the dashed line corresponds toi*2 (t). Performance of this strategy is superior to the process-the-targets-ignored-the-longest strat-egy since inter-target travel times are used by the scheduler.

0

10

20

30

0 50 100 0 50 100 0 50 1000

5

10

t, s

T1

0

10

20

30

T4

0

10

20

30

T7

0

10

20

30

T8

0

10

20

30

T5

0

10

20

30

T6

0

10

20

30

T2

0

10

20

30

T3

i 1(t

) ,i 2

(t)

**

from one computer to the other by means of an RS-232 link.The master computer, which acquires data from the slave,is in charge of the control strategies. Second, there is a sig-nificant calibration problem. We chose 16 LM35CAZ sensorsfrom National Semiconductor, which have essentially identi-cal characteristics. However, soldering the sensors to theboard might change the sensor characteristics in anunknown way. Although sensor calibration after construc-tion of the experiment is possible, it is tedious, time con-suming, and error prone. We used a Fluke Meter 179temperature probe to provide the ambient temperature atthe start of the experiment since it provides a measure forcomparing the achieved temperatures in two successiveruns of the experiment.

The cost of the sensors and actuators is less than$2 per zone. Supporting electronics includes inexpen-sive drivers for the lamps. A single-zone version of theexperiment provides students with experience usingthe dSPACE system. An eight-zone version can be usedin conjunction with a single dSPACE DS1104 board,and a communication network can be simulated in thatcase. Students are able to construct the experimentthemselves due to the simplicity and low cost.

The main challenges of this experiment are as follows:● temperature regulation uniformly across the grid but

with a fixed maximally elevated value; alternatively,we can seek a fixed or dynamic two-dimensional pat-tern of temperature values

● temperature tracking in which one set of zonestracks the average temperature in another zone, or areference temperature value

● distributed control with different controllers for eachzone, and a communication network with random butbounded delays for neighbor-zone sensed information,and delayed control inputs.

Distributed Control Strategies and ResultsIn the following, we show results for centralized and dis-tributed resource allocation strategies. It is useful, how-ever, to highlight several other control strategies thatcan be studied for this experiment. For instance, we haveimplemented a temperature tracking system with the 16zones partitioned into four superzones. One goal is toforce each superzone to track desired temperaturesTds

1 , . . . , Tds4 , that are set by the user. Another challenge

is the requirement that superzones 3 and 4 track the cur-rent average temperatures of superzones 2 and 1,respectively, when the user sets the desired tempera-tures for superzones 2 and 1.

Suppose that the control objective is to reach a maxi-mally elevated common temperature for all of the zones,but given the constraint that only one light can be on ata time. For this problem, we first consider a centralizedresource allocation strategy. Let the temperatures in theith zone be denoted by Ti and the ith heater input by ui.At each time instant our strategy applies a 100 ms dura-tion pulse to the zone that has the lowest temperature.This strategy allocates a resource defined by the timethat the light is on or, equivalently, the electrical cur-rent available for heating. In other words, at each time kand for each zone i,


Figure 8. Planar temperature grid control experiment. The experiment has 16 individually controlled zones, each with alight for a heater and an analog temperature sensor.

Light (Heater)

Sensor

Zone 4

Zone 1Zone 13

Heater Off

Heater On

ui(k) ={

1, if Ti(k) ≤ Tj(k) for all j ∈ {1, 2, . . . , 16},0, otherwise.

(2)

Intuitively, this strategy seeks to iteratively increase theminimum temperature, thereby forcing temperature uni-formity while maximizing all of the temperatures. Resourceallocation emerges from the simple optimization strategyimplemented by (2). Notice the conceptual similarity tothe juggler strategy for the balls-in-tubes experiment.

For this strategy, the whole grid reaches a practicallyuniform final temperature value as shown in Figure 9. Westart with a room temperature of 22.8 ◦C, and the averagefinal value is about 24.1 ◦C. Some of the zones are abovethis value; the maximum temperature is 24.8 ◦C. However,considering the disturbances associated with the experi-ment, the performance is quite good relative to all otherstrategies we have studied for this experiment. Since theroom temperature is variable, each time we run an experi-ment, the final average steady state value is different.

Another strategy is a distributed version of the previ-ous strategy. In particular, we assume that there is a localcontroller for each zone that can make decisions aboutwhether to turn its heater on or off, but based only on the

temperature in its own zone and possibly those in someadjacent zones. However, the local controllers can makedecisions using only temperature data obtained from adja-cent zones by way of a communication network that hasrandom but bounded delays. To define the strategy, letN(i) denote the set of neighbors of zone i, including zone i,from which the controller for zone i is given temperaturedata. For all i, let

Tmini (k) = min {Tj(k) : j ∈ N(i)}

and

ui(k) ={

1, Ti(k) = Tmini (k),

0, otherwise.

Intuitively, this control law is a distributed version ofthe strategy in (2). More than one zone can have itsheater on at a time, and hence the strategy can simultane-ously increase multiple temperature values. For an ambi-ent temperature of 21.8 ◦C, we obtain the results shownin Figure 10. After 200 s, the temperature in some of the


Figure 9. Temperature values for a centralized resource allocation experiment. This experiment was performed with a 22.8°C ambient temperature. After 200 s the average final value for the temperature is 24.1 °C.

22

23

24

25

T1

22

23

24

25

T2

22

23

24

25

T3

22

23

24

25

T4

22

23

24

25

T8

22

23

24

25

T12

22

23

24

25

T16

22

23

24

25

T7

22

23

24

25

T11

22

23

24

25

T15

22

23

24

25

T6

22

23

24

25

T10

22

23

24

25

T14

22

23

24

25

T5

22

23

24

25

T9

22

23

24

25

T13

0 100 200 0 100 200 0 100 200 0 100 200

t, s

zones is between 26–27◦C. Thereare other zones on the edges thatcannot reach this value because ofthe large disturbances. On theother hand, there are severalzones in the middle of the grid,such as T6 , that have relativelylarge temperature values. Theselarge values result from the lampson the edge zones being on practi-cally all the time. Moreover, notethat relative to the centralizedresource allocation case in Figure9, there are more oscillations inthe temperature values. Note thatdifferent vertical scales are usedin the two cases to facilitate view-ing the data. The oscillations aredue to the effects of the limitedamount of information used byeach local controller in the distrib-uted controller.


Figure 10. Temperature values for a distributed resource allocation experiment. Using a sensing topology based on gather-ing sensed temperature data from neighboring zones only, and starting from 21.8 °C, the final temperatures are between 26 °Cand 27 °C. Some of the zones on the edges cannot reach this value because of the large disturbance effects. The controller’sattempts to overcome such disturbances amplify the grid’s temperature differences by heating the middle zones while trying toraise the edge zones’ temperatures.

20

22

24

26

28

0 100 200 0 100 200 0 100 200 0 100 200

t, s

T1

20

22

24

26

28

T2

20

22

24

26

28

T3

20

22

24

26

28

T4

20

22

24

26

28

T8

20

22

24

26

28

T12

20

22

24

26

28

T16

20

22

24

26

28

T7

20

22

24

26

28

T11

20

22

24

26

28

T15

20

22

24

26

28

T6

20

22

24

26

28

T10

20

22

24

26

28

T14

20

22

24

26

28

T5

20

22

24

26

28

T9

20

22

24

26

28

T13

Figure 11. Building temperature control experiment. This experiment is a two-floormodel with four rooms and hallways on each floor. Each room has a temperature sen-sor and a heater. The model building is 39 cm wide, 46 cm deep, and 22 cm high.

DoorHallway

Room

Electrically Controlled Fan

SecondFloor

FirstFloor

Door

Building Temperature ControlWe constructed the scale-model building in Figure 11 tostudy multizone temperature control. The building temper-ature control experiment uses a two-floor model with fourrooms on each floor. Each of the rooms has a temperaturesensor and a heater. The two floors are separable so thatcoupling between floors can be considered or eliminated.Each floor has windows and doors with adjustable open-ings, allowing the interaction pattern between the temper-atures in the rooms to be adjusted. Moreover, there areseveral electronically controllable fans that can be insert-ed into windows or doors to simulate ambient wind effectsand drafts. These fans speed up the temperature interac-tion effects. The cost of the experiment is dominated bythe power supply.

The main challenge for this experiment is multivariabletemperature tracking. We can also consider challengesanalogous to the resource allocation objectives for the pla-nar grid. For more information, see “Additional Resources.”

Finally, we highlight alternative strategies that can beuseful for the temperature control problems. First, distrib-uted temperature control is used in the semiconductorprocessing area. For instance, achieving a uniform temper-ature on a plate is studied in [24]. Distributed control ofwafer temperature for multizone rapid thermal processingsystems is studied in [25]. In [26], the authors describe amultizone space heating system that maintains a desiredtemperature in different zones. Accurate multizone tem-perature control is of significant importance in personalcomputers, and several current strategies are described in[27]. Another approach for distributed control design witha tracking objective is described in [28], where the authorsshow how systems with a spatial interconnection topologycan be controlled using conditions that can be expressedas linear matrix inequalities.

Concluding RemarksWe introduced several inexpensive experiments for study-ing the design and implementation of distributed and net-worked dynamic resource allocation, scheduling, andcontrol strategies. In each case, we presented an overviewof the experimental apparatus, the challenges involved,and, for three of the testbeds, showed experimental resultsfor distributed decision-making strategies.

Due to their modular nature, the experiments are easi-ly expandable. Adding more tubes and a reconfigurableinput manifold can provide interesting new challenges forthe balls-in-tubes experiment. With a grid of tubes, theexperiment’s operation will be more visually appealingsince it will present a sheet of balls that can dynamicallydeform in three dimensions. An easily reconfigured inputmanifold can provide opportunities for studying theeffect of plant configuration changes on allocation perfor-mance. The electromechanical arcade can be expanded in

many ways. For example, it can be extended to two orthree dimensions with two opposing teams. The planargrid for multizone temperature control was chosen to beof minimum size, but still interesting since it has edgeeffects and interior zones. It is of interest to make the gridlarger or in a different shape, since different disturbancepatterns will arise. Significant expansion of the experi-ments will increase the data acquisition requirements sothat our software/hardware choices might need to bereevaluated. One potential solution is the emerging net-worked embedded systems approach, as described in“Additional Resources.”

Overall, these ideas for experiment redesign high-light the existence of tradeoffs among experimental appara-tus design, the choice of hardware and software, andresearch and educational objectives. It is hoped that thisarticle will help educators design laboratories that supportthe study of information technology–enabled distributedfeedback control.

AcknowledgmentsThis work was supported by the Intelligent Systems Divi-sion of the NIST. The inputs of James Albus, Fred Proctor,and William Shackleford at NIST are gratefully acknowl-edged, as are the helpful suggestions of the reviewers. Anumber of other people contributed to the construction ofthe experiments that we describe here, and we gratefullyacknowledge their assistance. First, we would like to thankWilliam Thalgott and Emmet Evans for building the balls-in-tubes experiment and helping with construction detailsfor the electromechanical arcade and multizone tempera-ture control experiments. The balls-in-tubes experimentand the arcade were constructed by groups of undergradu-ates in an EE 682P design project. The arcade was laterimproved by Kristina Guspan and Dawn M. Kin. The build-ing temperature control experiment was initially construct-ed by a group of undergraduates in an EE 682P designproject and significantly improved by Todd Broceus andTyson Rathburn. Jorge Finke added the electrically con-trollable fans, interface cables, and implemented a con-troller in dSPACE.

References[1] T. Samad and G. Balas, Eds., Software-Enabled Control: InformationTechnology for Dynamical Systems. Piscataway, NJ: IEEE Press, 2003.

[2] OSU Cooperative Control Laboratory [Online]. Available:http://www.ece.osu.edu/ ∼umit/ee757.htm

[3] Carnegie Mellon Robotics Institute [Online]. Available:http://www.ri.cmu.edu/lab_ lists/index.html

[4] Crossbow Wireless Sensor Networks [Online]. Available:http://www.xbow.com/Products/Wireless_Sensor_Networks.htm

[5] R. M. Murray, “Future directions in control, dynamics and systems:Overview, grand challenges and new courses,” Eur. J. Contr., vol. 9, no.2, pp. 144–158, 2003.


[6] dSPACE [Online]. Available: http://www.dspace.de/ww/en/pub/home.htm

[7] Quanser Consulting [Online]. Available: http://www.quanser.com

[8] OSU Control Systems Implementation Laboratory, EE 758 [Online].Available: http://www.ece.osu.edu/∼passino/ee758.html

[9] NIST RCS [Online]. Available: http://www.ece.osu.edu/∼passino/RCSweb/

[10] National Instruments and LabVIEW [Online]. Available:http//www.ni.com/labview/

[11] M.P.J. Fromherz and W.B. Jackson, “Force allocation in a large-scale distributed active surface,’’ IEEE Trans. Contr. Syst. Technol., vol.11, no. 5, pp. 641–655, 2003.

[12] Mathworks [Online]. Available: http://www.mathworks.com/prod-ucts/simulink/

[13] V. Gazi, M. Moore, K. Passino, W. Shackleford, F. Proctor, and J. Albus, The RCS Handbook: Tools for Real Time Control Systems Soft-ware Development. New York: Wiley, 2000.

[14] M. Moore, V. Gazi, K. Passino, W. Shackleford, and F. Proctor,“Design and implementation of complex control systems using the{NIST-RCS} software library,’’ IEEE Control Syst. Mag., vol. 19, no. 3, pp.12–28, 1999.

[15] A.N. Michel and R.K. Miller, Qualitative Analysis of Large ScaleDynamical Systems. New York: Academic, 1977.

[16] D.D. Siljak, Large-Scale Dynamic Systems Stability and Structure.New York: Elsevier, 1978.

[17] T. Ibaraki and N. Katoh, Resource Allocation Problems: AlgorithmicApproaches. Cambridge, MA: MIT Press, 1988.

[18] K. Passino, Biomimicry for Optimization, Control, and Automation.London: Springer-Verlag, 2004.

[19] A.E. Gil and K.M. Passino, “Stability analysis of network-basedcooperative resource allocation strategies,’’ in Proc. IEEE Conf. Deci-sion Control, 2003, pp. 1206–1211.

[20] J. Perkins and P. Kumar, “Stable, distributed, real-time schedulingof flexible manufacturing/assembly/disassembly systems,’’ IEEE Trans.Automat. Contr., vol. 34, no. 2, pp. 139–148, 1989.

[21] A.E. Gil, K.M. Passino, S. Ganapathy, and A.G. Sparks, “Cooperativescheduling of tasks for networked uninhabited autonomous vehicles,’’in Proc. IEEE Conf. Decision Control, 2003, pp. 522–527.

[22] M. Pinedo, Scheduling: Theory, Algorithms, and Systems. EnglewoodCliffs, NJ: Prentice-Hall, 2002.

[23] S.B. Gershwin, Manufacturing System Engineering. Englewood Cliffs,NJ: Prentice-Hall, 1994.

[24] C.D. Schaper, K. El-Awady, and A. Tay, “Spatially-programmabletemperature control and measurement for chemically amplified pho-toresist processing,’’ in SPIE Conf. Process, Equipment, Materials ControlIntegrated Circuit Manufacturing V, 1999, vol. 3882, pp. 74–79.

[25] C.D. Schaper, T. Kailath, and Y.J. Lee, “Decentralized control ofwafer temperature for multizone rapid thermal processing systems,’’IEEE Trans. Semiconduct. Manufact., vol. 12, no. 2, pp. 193–199, 1999.

[26] M. Zaheer-uddin, R.V. Patel, and S.A.K. Al-Assadi, “Design ofdecentralized robust controllers for multizone space heating sys-tems,’’ IEEE Trans. Contr. Syst. Technol., vol. 1, no. 4, pp. 246–261, 1993.

[27] P.E. Ross, “Beat the heat,’’ IEEE Spectr., vol. 41, no. 5, pp. 38–43,2004.

[28] R. D’Andrea and G.E. Dullerud, “Distributed control design for spa-tially interconnected systems,’’ IEEE Trans. Automat. Contr., vol. 48, no.9, pp. 1478–1495, 2003.

Nicanor Quijano received his B.S. degree in electronicalengineering from Pontificia Universidad Javeriana (PUJ),Bogota, Colombia, in 1999. In 2002, he received his M.S.degree in electrical engineering from the Ohio State Univer-sity, where he is currently working toward the Ph.D.degree in electrical engineering. From 1999 to 2000, he wasan instructor at PUJ. He has also worked for BPX Colom-bia, and Capitel, Telecom. His research interests includehierarchical and distributed methods, dynamic resourceallocation, and evolutionary game theory.

Alvaro E. Gil received his B.S. and M.S. degrees in electri-cal engineering from Instituto Universitario Politecnico, Barquisimeto, Venezuela, in 1990 and 1998, respectively,and the Ph.D. degree in electrical engineering from the OhioState University, Columbus, in 2003. He is currently a post-doctoral researcher at Ohio State University. From 1990 to1999, he held engineering positions in the AutomationDepartment at Petroleos de Venezuela in Maracaibo,Venezuela. From 1999 to 2002, he was supported by FUN-DAYACUCHO and PDVSA, and from 2002 to 2003 he was aresearch associate at the Department of Electrical Engineer-ing, the Ohio State University. His current research inter-ests include stability analysis of networked cooperativeresource allocation strategies.

Kevin M. Passino ([email protected]) received hisPh.D. in electrical engineering from the University of NotreDame in 1989. He is currently a professor of electrical andcomputer engineering at the Ohio State University (OSU)and director of the OSU Collaborative Center of ControlScience. He has served the IEEE Control Systems Societyin many roles: as the vice president of technical activities,an elected member of the Board of Governors, Programchair of the 2001 IEEE Conference on Decision and Con-trol, and a Distinguished Lecturer. He is coeditor of AnIntroduction to Intelligent and Autonomous Control (1993),coauthor of Fuzzy Control (1998), coauthor of StabilityAnalysis of Discrete Event Systems (1998), coauthor of TheRCS Handbook: Tools for Real Time Control Systems Soft-ware Development (2001), coauthor of Stable AdaptiveControl and Estimation for Nonlinear Systems: Neural andFuzzy Approximator Techniques, (2002), and author of Bio-mimicry for Optimization, Control, and Automation (2004).He can be contacted at the Department of Electrical andComputer Engineering, The Ohio State University, 2015Neil Avenue, Columbus, OH 43210 USA.


experiments for dynamic resource allocation, …passino/paperstopost/raschedcontrol-cs… ·...

Documents