Construction Management Applications - Challenges in Developing Execution Control Plans

download Construction Management Applications - Challenges in Developing Execution Control Plans

of 4

Transcript of Construction Management Applications - Challenges in Developing Execution Control Plans

  • 8/10/2019 Construction Management Applications - Challenges in Developing Execution Control Plans

    1/4

    Construction Management Applications:Challenges in Developing Execution Control Plans

    Nilufer Onder

    Dept. of Computer ScienceMichigan Technological University

    Houghton, MI [email protected]

    Amlan Mukherjee

    Dept. of Civil and Env. EngineeringMichigan Technological University

    Houghton, MI [email protected]

    Pei Tang

    Dept. of Civil and Env. EngineeringMichigan Technological University

    Houghton, MI [email protected]

    Abstract

    The objective of automated planners is to synthesizesequences of actions (called policies in MDP frame-works) that will achieve a predetermined goal given afully or partially observable formal representation of thedomain. In contrast, the main characteristic of projectmanagement is the greater emphasis on plan execution

    under uncertainty as opposed to plan synthesis. This pa-per explains the need to transition from automated plansynthesis to plan management and identifies the chal-lenges for the planning and scheduling communities us-ing examples of construction projects.

    Introduction

    Over the past decade, following the development of power-ful solutions to deterministic planning problems (Nau 2007)and the presentation of the challenges in stochastic domains(Bresina et al. 2002), the planning community has placedspecial emphasis on research in planning under uncertainty.Recent frameworks can deal with an overwhelming numberof states or plans (Buffet and Aberdeen 2009), and can uti-lize complex domain models such as continuous phase dis-tributions in generalized semi-MDPs (Younes 2005). Theobjective of automated planners is to synthesize sequencesof actions (called policies in MDP frameworks) that willachieve a predetermined goal given a fully or partially ob-servable formal representation of the domain. In contrast,the main characteristic of project management is the greateremphasis on plan execution as opposed to plan synthesis.The life cycle of a construction project begins with thepreparation of an initial resource loaded activity schedule- or the As Planned Schedule (APS), and continues underthe governance of the project design, setup, constraints, andthe changes that occur during the execution of the project.The challenging problem for construction managers and theplanning community is to develop plans that support the ex-ecution of an APS under the occurrence of internal feedbackloops and external stochastic events.

    While establishing the APS is a non trivial problem, weare not presenting it as a challenge because one can be puttogether using knowledge from similar previous projects.

    Copyright c 2010, Association for the Advancement of ArtificialIntelligence (www.aaai.org). All rights reserved.

    However, actions in an APS usually do not have well definedprecondition or effect sets. Even when these sets are well-defined, they might be infinite. As a result, uncertainty doesnot manifest itself as a result of action effects, it stems frominherent variability in project parameters (such as variablecrew production rates), unexpected external events (such asbad weather), and emergent phenomena that unexpectedlyevolve through complex interactions between resources andproject constraints. Formally speaking, the problem ceasesto be a search for a plan that constitutes a valid sequence ofagent actions, and instead becomes a dynamic control prob-lem based on the satisfaction of constraints expressed in theAPS. Consequently, despite various developments in plan-ning theory, full automation of project management is still achallenging goal. The three main challenges are (1) to de-velop a knowledge representation scheme to represent theproject, its constraints, and the sources of uncertainty; (2)to develop techniques for computing the execution time de-viations from the APS; and (3) to control the project whileminimizing the damaging effects of the deviations.

    This paper explains the need to transition from automated

    plan synthesis to plan management and identifies the chal-lenges for the planning and scheduling communities usingexamples of construction projects.

    Background in Construction Management

    The New York Times building is an interesting examplethat shows how a project can go in unexpected directionsand how adhering to the initial plan is challenging. TheNew York Times is a 52-story office building which was re-quested to be ultra-transparent by the owner (Post 2007).The building features a light-gray, horizontal filigree sun-screen that masks the 650,000-sq-ft facade to the outdoor-steel exoskeleton and terra cotta sunscreen rods for energyefficiency. Even though the planning process to develop

    the APS spanned three years and significantly lowered ex-pected costs using careful computer simulations, the con-struction projects execution was plagued with various prob-lems. Soon after the job started, the steel fabricator closeddown and the steel erector went bankrupt. The higher pricegiven by the new contractor cost the construction managerstheir entire contingency funds. This is an example of anunforeseen external event that they could not have budgetedfor. Moreover, the aesthetic considerations of the transparent

    263

    Proceedings of the Twentieth International Conference on Automated Planning and Scheduling (ICAPS 2010)

  • 8/10/2019 Construction Management Applications - Challenges in Developing Execution Control Plans

    2/4

  • 8/10/2019 Construction Management Applications - Challenges in Developing Execution Control Plans

    3/4

    two time points and associate penalties with various degreesof delay. For example, for the hoisting activity of sequence1, the following constraint:

    0 : 1 H1,e H1,b 2

    k : 2 + k= H1,e H1,b

    says that it will incur a penalty if its duration exceeds 2 days.For a delay of k days, the activity incurs a penalty which

    is computed by a function ofk denoted by k. This delayrepresents extra costs due to longer equipment usage or in-creased labor hours. Examples of constraints considered inthe implementation were:

    Constraint 1: In order to maintain work-flow and effi-ciently utilize equipment, lag time between hoisting ac-tivities for sequence i and (i+ 1) should be 0. This isexpressed as: : Hi+1,b Hi,e > 0.

    Constraint 2: In order to maintain labor-flow and effi-ciently manage crew on site, the lag time between boltingactivities for sequence i and (i + 1) should be 0. This isexpressed as: : Bi+1,b Bi,e> 0.

    Constraint 3: In order to ensure safety, the hoisting op-eration for a sequence of a higher storey has to start af-ter a considerable portion of the decking operation for asequence of the immediately lower storey had been com-pleted to afford safe fall distances. Appropriate fall dis-tances are maintained by ensuring that the decking op-eration for operation i starts before the hoisting oper-ation for the sequence (i + 2). This is expressed as:: Hi+2,b Di,b> 0.

    It is important to note that these constraints are project spe-cific manifestations of construction domain constraints re-garding safety and optimal resource utilization.

    Disruptions in ICDMA are created when constraints areviolated. These violations can result from external events

    or emergent uncertainty inherent to the project. An externalevent is a happening that affects the project. An event E isrepresented as a triple < Eprec, Ee, Epr >. Eprecis a setof variable values that enable the occurrence of the event.Ee is the set of effects the event has on the environment.It encodes the impacts that the event will have on specificproject parameters, and the variables impacted. The set ofvariables affected by an event can be different from the vari-ables included in the preconditions. Eprrefers to the proba-bility of event occurrence given that the enabling precondi-tions are satisfied. For example, a drop in labor productivitydue to adverse weather can be represented as:

    Eprec ={Weather =Snow, Location =outdoor},E

    e =

    {Productivity = 50%

    },

    Epr ={0.75}

    At each time step in the simulation, ICDMA samples fromthe event probability distributions to determine which eventswill occur and what the effects will be. It uses a state basedrepresentation using a resource loaded temporal constraintnetwork, called TONAE ( Temporal Network with Activitiesand Events), which represents both temporal and resourceconstraints (Anderson, Mukherjee, and Onder 2009). At the

    Figure 1: ICDMA at t=0

    start of the project all constraints are satisfied and the APSis the state of the system at time t = 0- represented as S0

    (Fig 1). ICDMA uses continuous time advancement, whichis different from the DES frameworks advancing time fromone discrete event to the next. Thus each time point consti-tutes a control point in the simulation. This is critical be-cause stochastic external events are equally likely to happenat any point in the project.

    Currently TONAE does not represent the adaptive andemergent uncertainties that arise from spatio-temporal inter-actions between dynamic relationships among constructionlabor crews, decision outcomes, and equipment on site. Forexample, inappropriate positioning of a crane can jeopardizesafety and lead to lost productivity from congestion on site.Project specific causes that set such feedback loops in mo-tion, and balance or reinforce them when control measures

    are executed, must also be investigated. Therefore, a repre-sentational challenge is to explicitly represent emergent un-certainty and consider the influence of the dynamic evolu-tion of project constraints on the execution control plan.

    In order to meet the representational challenges (chal-lenge 1), all the constraints that might be relevant to theactivities in the APS must also be represented. We needa rich representation including temporal constraints, safetyconstraints, and resource constraints such as equipment andspace. The adaptive and emergent behaviors arising from in-teractions between individual construction components andsite peculiarities also must be modeled explicitly.

    The difference between the APS and the as simulated met-ric at time t is represented ast. ICDMA can compute the

    deviations in terms of remaining schedule duration and extracost incurred. It can create bins possible futures with respectto probabilities or costs but does not perform further analysis(challenge 2).

    ICDMA does not have mechanisms to develop responsesto project deviations instead decisions are obtained from theuser. A decision is an assignment or re-assignment of realnumbers to resource and schedule variables (e.g., if 5 moreunits of labor is hired then the relevant variable is updated).

    265

  • 8/10/2019 Construction Management Applications - Challenges in Developing Execution Control Plans

    4/4

    The decision set consists of all reassignments and is formu-lated as Dtwhen made at the end of time point t.

    At the end of each simulation run we generate the follow-ing data sets:

    A sequence of system states for each time point during thelength of the simulated project: {St}. The TONAE is thestate of the system at any given time.

    A sequence of decisions taken by the user: {Dt}.

    A sequence of deviations at each time step:{t}.

    A sequence of query distributions recorded at the end ofeach time point in the simulation:{Qt}. This is an impactdistribution of a sample of possible future states that Stcan transition to.

    Hence for each time point for a given run the following situ-ation vector can be recorded as: St = [St, Qt,t, Dt]. Thedecision-making agent arrives at the decision Dt by observ-ing St and Qt, with the objective of minimizing t. Thedecision Dtin conjunction with the stochastic occurrence ofan external event is used to update the state of TONAE totrigger the transition of the system from Stto St+1. The de-cision Dtmade by the agent is driven by high-level philoso-

    phies which we refer to as strategies. Examples of suchstrategies include a crash strategy where the user increasesresources to decrease project duration, and a reassign strat-egywhere the activities on the critical path are prioritized.

    The challenges related to generating contingency re-sponses (challenge 3) are as follows:

    Given a set ofk strategies or policies {P0, P1, P2...Pk},and N simulation runs with each strategy (total ofkNruns) and the sequences of situation vectors recordedfor each run {S1t ,S

    2t ...S

    kt}, how do we establish formal

    methodologies and approaches to represent and reasonabout the efficiency of a strategy?

    The decision-making agent can be a human or an arti-

    ficial agent. The advantages of using a human agent isthat each decision will take into consideration judgementand understanding of domain/problem specific knowl-edge. However, for large values ofN it becomes impracti-cal and error prone to have a human agent. Therefore, canan automated agent be devised that will (i) run the simu-lation given a particular strategy, (ii) possibly learn fromhuman agent runs to increase its planning efficiency?

    ConclusionIn this paper, we explained the need for knowledge-richenvironments to deal with execution time contingenciesin project management. While the examples and caseswere drawn from construction projects, we believe that the

    desiderata are widely applicable to other dynamic projectenvironments which have soft and hard constraints, variabil-ity in activity parameters, and external events affecting thecourse of the project.

    Good user interfaces are also essential for effective projectmanagement, but are beyond the scope of this paper. Oneinteresting issue is the development of powerful user inter-faces for entering information about stochastic domains. Re-cently, such interfaces for deterministic domains have been

    presented in the context of the International Competitionon Knowledge Engineering for Planning and Scheduling(Bartak, Fratini, and McCluskey 2009).

    Acknowledgments

    This work was supported by the NSF grant SES 0624118and its REU supplement to Amlan Mukherjee. Any opin-ions, findings and conclusions or recommendations ex-

    pressed in this material are those of the authors and do notnecessarily reflect views of the National Science Founda-tion. We thank anonymous reviewers for their helpful sug-gestions.

    References

    Anderson, G. R.; Mukherjee, A.; and Onder, N. 2009.Traversing and querying constraint driven temporal net-works to estimate construction contingencies. Automationin Construction18(6):798813.

    Bartak, R.; Fratini, S.; and McCluskey, L., eds. 2009. Pro-ceedings of the Third International Competition on Knowl-edge Engineering for Planning and Scheduling. AAAI

    Press.Bresina, J.; Dearden, R.; Meuleau, N.; Ramakrishnan, S.;Smith, D.; and Washington, R. 2002. Planning under con-tinuous time and resource uncertainty: A challenge for AI.InProceedings of the Eighteenth Conference on Uncertaintyin Artificial Intelligence (UAI-02), 7784.

    Buffet, O., and Aberdeen, D. 2009. The factored policy-gradient planner. Artificial Intelligence173:722747.

    Daccarett, V., and Mrozowski, T. 1999. AISC digital library:Construction management of steel construction. Avail-able via http://www.aisc.org/content.aspx?id=21250 [ac-cessed 10/26/2009].

    Hajjar, D., and AbouRizk, S. 2002. Unified modeling

    methodology for construction simulation. Journal of Con-struction Engineering and Management128(2):174185.

    Martinez, J., and Ioannou, P. 1999. General-purpose systemfor effective construction simulation. Journal of Construc-tion Engineering and Management125(4):265276.

    Nau, D. S. 2007. Current trends in automated planning. AIMagazine28(4):4358.

    Pena-Mora, F., and Li, M. 2001. Dynamic planning andcontrol methodology for design/build fast-track constructionprojects.Journal of Construction Engineering and Manage-ment127(1):117.

    Post, N. M. 2007. High-profile manhattan tower is veteranbuilders last hurrah. Engineering News Record.

    Watkins, M.; Mukherjee, A.; and Onder, N. 2008. Usingsituational simulations to collect and analyze dynamic con-struction management decision-making data. In Proceed-ings of the 40th Winter Simulation Conference (WSC-08),23772386.

    Younes, H. L. S. 2005. Planning and execution with phasetransitions. InProceedings of the Twentieth National Con-ference on Artificial Intelligence (AAAI-05), 10301035.

    266