Integration of Life Cycle Assessment Into Agent-Based Modelingre.indiaenvironmentportal.org.in/files/Integration of Life Cycle.pdf · Integration of Life Cycle Assessment Into Agent-Based

R E S E A R C H A N D A N A LYS I S

Integration of Life CycleAssessment Into Agent-BasedModelingToward Informed Decisions on EvolvingInfrastructure Systems

Chris Davis, Igor Nikolic, and Gerard P. J. Dijkema

Keywords:

bioenergycomplexityindustrial ecologymatrixnetworkontology

Summary

A method is presented that allows for a life cycle assessment(LCA) to provide environmental information on an energyinfrastructure system while it evolves. Energy conversion facili-ties are represented in an agent-based model (ABM) as distinctinstances of technologies with owners capable of making deci-sions based on economic and environmental information. Thissimulation setup allows us to explore the dynamics of assem-bly, disassembly, and use of these systems, which typically spandecades, and to analyze the effect of using LCA informationin decision making.

We were able to integrate a simplified LCA into an ABMby aligning and connecting the data structures that representthe energy infrastructure and the supply chains from sourceto sink. By using an appropriate database containing life cycleinventory (LCI) information and by solving the scaling factorsfor the technology matrix, we computed the contribution toglobal warming in terms of carbon dioxide (CO2) equivalentsin the form of a single impact indicator for each instance oftechnology at each discrete simulation step. These LCAs maythen serve to show each agent the impact of its activitiesat a global level, as indicated by its contribution to climatechange. Similar to economic indicators, the LCA indicatorsmay be fed back to the simulated decision making in the ABMto emulate the use of environmental information while thesystem evolves. A proof of concept was developed that isillustrated for a simplified LCA and ABM used to generateand simulate the evolution of a bioelectricity infrastructuresystem.

Address correspondence to:Chris DavisJaffalaan 52628 BX Delftthe [email protected]

c© 2009 by Yale UniversityDOI: 10.1111/j.1530-9290.2009.00122.x

Volume 13, Number 2

306 Journal of Industrial Ecology www.blackwellpublishing.com/jie


Introduction

Understanding issues related to sustainabilityoften involves recognizing that the world is com-posed of a mix of economic, environmental, andsocial systems. These systems are not generallymutually exclusive but are interconnected witheach other. For example, the economy is inher-ently dependent on the environment for raw re-sources, and it is also the result of a multitude ofsocial decisions.

Modeling has been useful as an exploratoryand analytical technique for understanding thesesystems. One of the challenges in dealing withsustainability, however, relates to the modelingof interconnected systems. Modeling these can bevery difficult, as it is not unusual for a change ordecision in one part of the system to have ripplingeffects elsewhere throughout the system. For ex-ample, it can be difficult to understand how amanagement decision on investment in a certaintype of technology could have environmental im-pacts due to economic conditions further up thesupply chain for that technology.

To approach these problems of system in-terconnectedness and interdependency, in thiswe examine whether and how the combinationof two types of tools can provide a step for-ward. One of these tools is life cycle assessment(LCA), which has become a standard tool forenvironmental analysis of our technological pro-duction systems. The other is agent-based mod-eling (ABM), a simulation tool that is usefulfor generating complex network systems that re-sult from the decision making of individual en-tities. LCA is very good at analyzing complexnetwork structures, although in its current form,it is a tool for linear modeling of static systems.Furthermore, traditional LCA does not examineeconomic and social concerns (Guinee 2002).Conversely, ABM provides a means to createnonlinear dynamic systems, which can be speci-fied to include social and economic aspects.

We hypothesize that combining these toolscan result in an interesting synergy. In this ar-ticle, we present a proof of concept that thisintegration is possible and provides a means formodeling interdependent sociotechnical systems.Our intent is to show that this combination canprovide additional information about sociotech-

nical systems to aid in decision making, ratherthan specifying the exact decision that should bemade. It is still up to the decision maker to use hisor her own judgment in interpreting the differentaspects of this information.

The modeling framework developed has beenused in a case study investigating bioelectricityproduction in the Netherlands. This is a topicof much interest, because using biomass as a fuelis assumed to offset fossil carbon dioxide (CO2)emissions. Because CO2 emissions are contribut-ing to a global problem, however, one needs toexamine the biomass supply chains to quantifythe gains or even losses that may be occurring.The agents in the simulation were defined on thebasis of technologies in these supply chains.

The sections below examine the foundationsfor LCA and ABM, the methods used for inte-grating these two tools, and present then a proofof principle demonstrating repeated accounting-type LCAs for a dynamic supply network. This isfollowed by a discussion of insights gained, alongwith the future outlook and conclusions.

Foundations

Both LCA and ABM are tools that employsystems approaches. That is, both of them areused to investigate systems with multiple inter-acting elements. Simply put, they both examine“things interacting with other things.” LCA looksat interacting technologies in the form of supplychains, whereas ABM is a more general tool thatlooks at the interactions of a group of agents. Al-though they are different tools, they are not fun-damentally dissimilar in their approaches to rep-resenting systems. LCA stores the interactions oftechnologies in a matrix format. An agent-basedmodel can be set up to represent interactions in anetwork format, where agents are represented bynodes and edges represent connections betweenthe agents. From this insight, one of the keys tointegrating these tools was the realization that thenetwork data structure of the agent-based modelis equivalent to the matrix data structure of theLCA, which allows the modeled system to berepresented in both formats.

One large difference between LCA and ABMis that within an LCA, the connections betweentechnologies are fixed, whereas an agent-based

Davis et al., Integration of LCA Into Agent-Based Modeling 307


model does not have to define fixed connections.Within an LCA, we know exactly which tech-nologies use the output from another technology.Within an agent-based model, however, agentscan be defined as having different options con-cerning those with whom they interact. Essen-tially, LCA has a static structure, whereas ABMcan allow for changing and evolving structures.

On the basis of this, we suggest using ABM togenerate a complex dynamic system and an LCAto analyze it at discrete simulation time steps.Instead of defining the world as a set of tech-nologies with fixed connections, this approachbreaks those connections. Technologies are nowrepresented as agents whose owners must tradewith other agents to buy inputs and sell theiroutputs. Supply chains within the simulation self-assemble due to these interactions between indi-viduals. By appropriate definition of agents’ de-cision rules and identity and by inclusion of asufficient number of actors in the simulation, thisconcept can be extended to emulate socioeco-nomic decision making and yield dynamic mod-els of supply networks. By extension, ABM couldultimately be used to model whole economies orsocieties.

The next two subsections give a brief overviewof LCA and ABM to illustrate their approaches,limitations, and opportunities for integration.

Life Cycle Assessment

LCA is a methodology and tool used to an-alyze the environmental burden of productsat all stages in their life cycle—from the ex-traction of resources, through the productionof materials, product parts and the productitself, and the use of the product to the man-agement after it is discarded, either by reuse,recycling or final disposal. (Guinee 2002, 5)

This comes from a recognition that environ-mental impacts have a systemic origin. In otherwords, by choosing one particular product or ser-vice, we are indirectly supporting environmentalimpacts that may occur several stages away fromus in the supply chain that brings a product intobeing or provides a service.

Today’s production networks have emergedthrough the accumulation of choices made by amultitude of actors driven by a variety of ratio-

nales. LCA provides an accounting of the totalenvironmental impact of this network and relatesit to a particular product or service. Although itthus may not be immediately apparent, LCA is atits heart a type of network metric or a means ofcalculating characteristics of a network. In par-ticular, it is a way of understanding a network’sstructure and performance from the viewpoint ofone node within a network that represents thefunctional flow or reference product.

LCA has been a valuable tool in helping usholistically approach the problem of emissionsreductions. It is not enough to reduce emissions ata local level; we must be aware of emissions thatoccur beyond our horizon as a consequence ofour actions. We cannot fully achieve systematicemissions reductions without recognizing thesenetworks of interdependence between industries.

The idea behind LCA is very powerful, al-though in its current implementation, we believeits potential is not fully utilized. By connectingan LCA to an agent-based model, we anticipatethat we can overcome many limitations.

Currently, LCA views the world as being com-posed of static connections between technolo-gies. It is also a linear model in that if productionof one good is increased, the flows of the up-stream technologies are scaled proportionally aswell. Additionally, it does not include any timeor location aspects, which means that impact as-sessment for certain emissions may be too high,especially if those emissions are spread out widelyover different time periods and regions (Guinee2002).

Agent-based Modeling

The concept of representing elements of a sys-tem as individual agents is central to ABM. AsABM is used in very diverse scientific fields, how-ever, it has no exact definition. One explanationis given by Shalizi (2006, 35), who states that

an agent is a persistent thing which has somestate we find worth representing, and whichinteracts with other agents, mutually modify-ing each others’ states. The components of anagent-based model are a collection of agentsand their states, the rules governing the inter-actions of the agents, and the environmentwithin which they live.

308 Journal of Industrial Ecology


Kauffman (Ball 1999 as cited in Shalizi 2006,35) provides a much more succinct definition bystating, “An agent is a thing that does things toother things.”

In this sense, agents serve as a flexible abstrac-tion of the real world. The key idea is that systembehavior is controlled not by a single entity butby the accumulated effects of decisions made bythe individual agents. When these agents inter-act with each other, they form networks, whichcan then give rise to complex systems. For theagent-based model in this article, agents musttrade to buy raw materials and sell their products.Agents are only limited in that they must find asuitable trading partner who makes a feedstockthey need or is willing to buy what they produce.If an agent is not profitable, it will “die” and beremoved from the simulation. In the ABM simu-lation, the technology options available and thedecision-making rules specified for the agents arethe ingredients used to generate the overall sys-tem structure and content.

ABM is especially suited for types of problemsfor which equations either cannot be solved or areimpossible to formulate (Axtell 2000). Comparedto other tools, such as system dynamics, ABM al-lows the modeling of more complex dynamics, be-cause the system structure can also change duringthe simulation. In system dynamics, a fixed inter-action structure is defined and maintained, whichmeans that even before the simulation is started,one must define how the parts are connected.Using an ABM, one only defines an interactionspace in the form of the types of interactions thatthe agents are allowed to have. This character-istic of ABM is important, given that many ofthe largest problems we face today are difficult tosolve because they exhibit characteristics of com-plex adaptive systems (CASs). These have beendefined by Holland as

Figure 1 Example of sociotechnicalsystems modeled. Nodes in theTechnology Owners box arecompanies. CO2 = carbon dioxide;CH4 = methane; N2O = nitrousoxide.

a dynamic network of many agents (whichmay represent cells, species, individuals,firms, nations) acting in parallel, constantlyacting and reacting to what the other agentsare doing. The control of a CAS tends to behighly dispersed and decentralized. If there isto be any coherent behavior in the system,it has to arise from competition and coop-eration among the agents themselves. Theoverall behavior of the system is the result ofa huge number of decisions made every mo-ment by many individual agents (Waldorp1992, 145).

We use this concept of an agent as a basis formodeling of technological infrastructures. We ex-pand the concept by incorporating insights fromstudies on large-scale sociotechnical systems (Bi-jker et al. 1987; Nikolic et al. 2007, 2008). Com-panies are defined as agents who own and man-age technologies, as illustrated in figure 1. Theseagents must make decisions on how to operatetheir technologies and trade with other agents tosell their products and buy feedstocks. The agentsencounter real-world constraints in that to sur-vive, they must remain profitable. Technologiesowned by agents are constrained in that they mustbalance their mass, energy, and monetary flows.As the technologies operate, economic and en-vironmental performance data are gathered andthen fed back to the agents to influence theirmanagement decisions.

Integration of Life CycleAssessment into Agent-basedModeling

Both ABM and LCA deal with networkedstructures. Given a suitable system conceptual-ization, network metrics can be calculated. In anagent-based model, the network is represented asa graph consisting of nodes connected by edges.



Figure 2 Equivalence of graph and matrix implementations.

In an LCA, one conceptualizes the network as amatrix where nonzero values represent the mag-nitude of flows between nodes, whose identitiesare indicated by the row and column locations.These two concepts are equivalent, as shown infigure 2, where a graph is mapped to a matrix. Oneof the keys to combining ABM and LCA hasbeen to interchangeably represent the networkstructure in both forms. Graphs are very usefulfor keeping track of complex evolving structures,and algorithms can be written to navigate andretrieve relevant information from these struc-tures. The matrix representation of LCA is theonly computationally feasible means for calcula-tions involving large systems, however (Heijungsand Suh 2002).

Combining ABM and LCA requires that weexamine the data structures used by both. A typ-

Figure 3 Illustration of life cycle assessment data structure. The left matrix is for material flows betweentechnologies. The right matrix is for environmental interventions from technologies. Squares representnonzero matrix elements.

ical data structure for an LCA is illustrated infigure 3. Most of the data are stored in two ma-trices. The technology matrix has informationabout material flows between technologies, andthe emissions matrix has information about theamount of pollution produced by each technol-ogy. More detail on this is given in the Calcula-tions section below.

Ontologies

Figure 4 shows a sample portion of the datastructure used in the agent-based model. As men-tioned previously and illustrated in figure 2, theagent-based model employs a graph-based datastructure. This description is simplified in thatthe graph is called an ontology, which is a way oflogically structuring and storing information so



Figure 4 Partial illustration of the ontology used for agent-based modeling data structure.

that it can be understood by humans and imple-mented in and retrieved by executable computercode. An ontology often specifies “is a” and “hasa” relationships for objects. For instance, an appleis a fruit and has a red color. In particular, an on-tology is composed of classes, properties, and in-stances. A class can be used to describe a generictype or group that something belongs to, such asa fruit class. This fruit class can then have prop-erties describing attributes, such as name, shape,and color. This defines the basic data structurethrough which we can describe fruit. We can thencreate multiple instances, such as for “this apple”and “this pear,” using the template specified bythe class to indicate the specific name, shape, andcolor of this apple and this pear.

We use the ontology classes to describe thedata structure of the agent, which defines thetypes of information it can know about itself. Forinstance, all agents have an OperationalConfig-uration containing a list of inputs and outputs,represented as ComponentTuples, which holdinformation about the amount of goods neededfor manufacturing and the amount of goods pro-duced. The agents also have InEdges and Out-Edges, which are used for contracts for itemsbought from and sold to other agents. This com-mon data structure allows agents to communicate

with each other about their properties and aids inretrieval of information later on, whether by theagents themselves or for data-logging purposes.

The instances contained in the ontology serveas a database. When we create an agent, it getsits own data structure, which is partially filledin with information, such as its required inputsand outputs, construction costs, and fixed costs.As it operates during the simulation, it fills inother information, such as its profits and lists ofcontracts made.

Assembling the Network

Because the other agents have the same typeof data structure defined by the ontology, we cansee the contracts as facilitating the creation of thenetwork of actors trading among one another.The information contained in the contracts isused in calculation of the LCA, as described fur-ther below in the Technology and Emissions Ma-trices section.

For TechnologyAgents to sell their goods andprocure the necessary inputs, they must trade withother agents. When a TechnologyAgent needs aninput, it asks all other TechnologyAgents for a pos-sible contract. The agents offering goods are ableto set the price in the contract on the basis of



their own decision making. Once an agent thatneeds an input has the complete list of offeredcontracts, it will decide to accept the contractsthat it deems best.

To help keep track of all the mass flows thatone will need in calculating an LCA, the con-cept of a WorldMarket and an Environment agentis used. Essentially, these agents ensure that ev-erything in the simulation comes and goes fromsomewhere. The WorldMarket agent serves toconnect the TechnologyAgents to the outside mar-ket. Although TechnologyAgents can buy and sellamong themselves, they may find that no one issupplying the good they want or that no one isbuying what they produce. In these cases, theycan trade with the WorldMarket. The Environ-ment is the agent that collects all emissions fromthe agents and provides the same inputs as thereal environment does, such as air for processesinvolving combustion.

Constructing the LCA

To provide additional information for theLCA calculations, in this proof of concept weemploy a hybrid approach whereby the World-Market agent’s supply chains are defined on thebasis of structures contained in a life cycle inven-tory (LCI) database. This is illustrated in figure 5:The evolving system is represented by the agent-based model, which is linked with the appropri-ate processes and supply chains in the databasecontaining LCI information. Whenever an agentbuys from the WorldMarket, due to the connec-tion with the LCI database, we are thus able toretrieve the upstream environmental emissionsresulting from that purchase. This linkage resultsin a coupling of systems, so that we now havea very large economic network composed of acore of dynamic actors (the TechnologyAgents)surrounded by static actors (based on the tech-nologies described in the database). This allows asimulation in which we have a “lens” of dynamicbehavior and evolving structures surrounded bya static definition of the world, as depicted infigure 6, which shows how conceptually themodel structure changes over time. As indicated,the set of agents in the agent-based model can in-crease or decrease with time, whereas during thewhole simulated period we are able to trace each

Figure 5 Systems representation in the hybridmodel. Dynamic supply chains in the agent-basedmodel connect to static, predefined supply chains.

agent’s upstream flows through the structure de-fined in the LCI database and linked to the ABMdata structure.

Technology and Emissions Matrices

Figure 7 illustrates the data structure used tocombine LCA within ABM. The simulation ispaused at every time step, and all the informa-tion about inputs and outputs (i.e., traded flowsand emissions) is retrieved from the simulation’sdata structure. This information is then processedand placed into a technology and an emissionsmatrix, where calculations can begin. At thisstage, economic allocation occurs for technolo-gies with multiple outputs (multifunctional pro-cesses). The implementation of this particularallocation method is relatively easy, with the cal-culation performed automatically by the software,and the only user input is a set of initial pricesfor goods. With additional programming, othertypes of allocation could be used. For this proof-of-concept stage, however, economic allocationwas used, with the recognition that programmingfor other forms of allocation could be worthwhileextensions to the methodology and models in thefuture.

In both of the matrices, one section is com-posed of information from the external database



Figure 6 Hybrid model over time. Theagent-based model evolves, and predefined supplychains are used to provide additional informationabout environmental impacts that exist outside it.

containing LCI information. This portion staysconstant throughout the simulation. Additionalrows and columns are added to represent theagents themselves. Because part of the matrix iscomposed of a dynamic system, the matrix willgrow and shrink over time in accordance withthe number of active agents present at each timestep.

Figure 7 illustrates how the links are made be-tween different processes within the technologymatrix. The diagonal represents the total out-put of each process. By reading down a column,we can see the different amounts of inputs usedfor a single process. By reading across a row, wecan see how one process supplies several differ-ent processes. The top left quadrant is composedsolely of the LCI database, whereas the top right

quadrant contains information about where anagent buys goods from the WorldMarket, whichis connected to the LCI database. The bottomright quadrant shows trading between individualagents. The bottom left quadrant is not filled in,because the processes in the LCI database do notbuy goods from the agents and only use prede-fined supply chains. An emissions matrix is set upthat is similar, except that the columns representindividual processes, and rows represent types ofemissions.

Calculations

An LCA system can be written in the formshown in equation 1 (Heijungs and Suh 2002).Here, A is the technology matrix, which is alsoshown in figure 7. The vector y is the demandon the LCA system and represents the functionalunit. For instance, if one wanted to perform anLCA on 1 kilowatt hour (kWh) of electricityfrom a specific process, this vector would be allzeros, except for a 1 placed at the location cor-responding to the process of interest. The vec-tor s represents the scaling factors that one mustapply to the technology matrix to provide thefunctional unit. When we perform an LCA, thevalues for Aand y are already defined by the user,and we must solve for s, which results in equation(2).

As = y (1)

s = y A−1 (2)

Solving for s is a nontrivial exercise. Althoughequation (2) shows a simple, elegant formula forsolving a complicated system, there are quite afew implications behind this. Performing an LCAcalculation means that one must balance sup-ply and demand by appropriately scaling eachtechnology in the matrix. Because an LCA isdone over a supply chain, scaling one technologymeans that one must scale the connecting tech-nologies as well to ensure that the mass balanceis satisfied for the entire system. For complex in-terdependent supply chains that include feedbackloops between technologies, this turns into a verycomplicated operation that truly necessitates theuse of a matrix.



Figure 7 Creating the technology matrix for life cycle assessment (LCA) calculations. The figurecorresponds to A shown in equation (1).

Although a matrix makes these types of cal-culations computationally feasible, there are stillissues with the amount of time needed, espe-cially if one desires to perform multiple LCAson a dynamic system, as demonstrated in thisarticle. A major bottleneck lies in calculatingthe inverse of the technology matrix, as requiredby equation (2). In the worst case, most matrixinversion algorithms are said to run in O(n3)time, which means that as the size of the matrixincreases by n, the maximum calculation timeneeded to invert it increases at a rate of n3. Inother words, one may need a fraction of a secondto invert a technology matrix of 100 processes,whereas one may need 2 minutes for one involv-ing 1,000 processes (Heijungs and Suh 2002).

The iterative matrix inversion algorithm pro-posed by Peters (2006) was designed to overcomethis problem. This particular algorithm runs inCn time, where C is generally a very large con-stant. This means that algorithms running inO(n3) time have an advantage when used forsmall matrices, but this iterative algorithm has asignificant advantage in dealing with large ma-trices. In practice, this algorithm has been in-

valuable in the creation of the model, as it is notuncommon to conduct dozens of simplified LCAsduring each simulation time step.

Within the simulation, a simplified LCA isgenerated for every agent during every time step.The functional unit is chosen on the basis of asingle unit of the main flow from the agent andcurrently remains the same throughout the sim-ulation. In other words, power plants will have afunctional unit of 1 kWh of electricity, whereasanother agent that makes palm oil will have afunctional unit of 1 kg of palm oil. We performeda simplified LCA to evaluate climate change onthe basis of emissions of CO2, methane (CH4),and nitrous oxide (N2O). We used the charac-terization factors of the LCA methodology toestimate the global warming potential (GWP)in terms of CO2-equivalents. The models couldreadily be extended to more extensive LCAsconsidering the full LCI and more impact cat-egories, and these are considered areas of futureextension of the methodology. At this proof-of-concept stage, however, the models were limitedto this simplified LCA. Only contribution to cli-mate change was evaluated, as it was considered



to be the most important impact category for thiscase study, which focuses on energy provision.

Overview of Steps

To summarize, the list below gives an overviewof the steps by which LCA was integrated into theagent-based model, as described in the paragraphsabove. A more complete description is given inthe original work (Davis 2007).

1. Run the simulation for one time step, al-lowing the agents to trade with each other.

2. Construct the technology and emissionsmatrices for each agent.(a) Collect information on all the flows

in the system.(b) If an agent trades with the World-

Market, link it to the supply chaindescribed in the LCI database.

(c) Perform economic allocation if nec-essary.

3. Calculate an LCA for each agent.(a) Determine the functional unit to be

used for each agent, on the basis of themain product produced. An agent’sfunctional unit is presently constant,as agents currently do not producedifferent products over different timesteps.

(b) Calculate an LCA for each agent onthe basis of functional unit.

(c) Aggregate CO2, CH4, and N2O intoa single category on the basis of theirGWP values.

4. Send results from the LCA to the agent foruse in decision making.

5. Run the simulation for the next time step,and repeat.

As has been shown, it is possible to use asimplified LCA within an agent-based model.This opens up new realms when we considerthat sustainability is an emergent property of anetwork (Allenby 1999), which LCA, to an ex-tent, inherently recognizes. Sustainability can-not be measured at the level of individuals; itcan only be measured at the level of the totalsystem itself. The value of this setup is that itallows us to create a dynamic model that can ex-plore the possible effects of independent as well

as interconnected decisions. This can allow fora dynamic representation of the system, includ-ing feedback loops and other dynamic systemproperties.

Proof-of-Principle Illustration

Bioelectricity

The types of agents used in this case studywere based on inventory data gathered duringa separate study investigating bioelectricity pro-duction in the Netherlands (Van der Voet et al.2008). This study was performed to create a cal-culation tool to allow for comparison of differ-ent bioelectricity supply chains with regard totheir greenhouse emissions. The greenhouse gasaccounting was limited to CO2, N2O, and CH4.Several different electricity production methodsappropriate to the Dutch situation were chosen,in addition to the types of biomass that wouldlikely be used to supply them. Inventory datawere collected for all stages, from biomass pro-duction to transportation, processing, and finalconversion to electricity. Several types of biomasswere chosen from different parts of the world, andthe electricity production methods ranged fromsmall scale (10 MW) to large scale (500 MW).From this information, an LCA was performedof the different chains, which allowed for com-parison of the greenhouse gas emissions resultingfrom 1 kWh of electricity production from eachof these different possible supply chains. Otheraspects, such as land use change, biodiversityimpacts, and economic development, were notinvestigated.

This study by Van der Voet and colleagues(2008) was, in turn, part of a larger investiga-tion commissioned by the Dutch government toexplore issues related to the use of biomass for en-ergy and material production. This was performedby the Sustainable Production of Biomass ProjectGroup (Commission Cramer), which investi-gated criteria for the evaluation of sustainableproduction of biomass (Projectgroep Duurzameproductie van biomassa 2006).

We give a brief explanation here to illustratethe system and the nature of the problem thatwas modeled.



Initial Settings

Each simulation starts with three large fossil-based power plants. These technologies are coalcofired with biomass, coal cofired with syngas,and natural gas and heavy oil cofired with bio-oil.Each of these large technologies is able to cofirebiomass at a fixed ratio, in addition to being ableto use pure fossil fuels.

The system is driven by the demand for elec-tricity, which has an effect on the number ofagents added. The actual initial demand for elec-tricity is set slightly above the production capa-bilities for the three main fossil electricity plants,which allows for additional, smaller producers tojoin in to fill the gap.

Agent Addition and Removal

At the beginning of each simulation time step,new TechnologyAgents have the opportunity tojoin the simulation. During this period, candidatetypes of TechnologyAgents are picked at randomand asked whether they want to invest. Once 10in a row have declined, the simulation continues,as it is considered unlikely that any others maywant to invest.

A candidate TechnologyAgent’s investmentdecision is based on several factors. First, thedemand-to-supply ratio for its reference productmust be higher than a specified value to avoidoversaturating the market. If this condition ismet, then the candidate will evaluate whetherit can be profitable under current market con-ditions by calculating its expected revenue fromthe sale of its products and then subtracting itsfixed operating costs and variable costs of inputs.If this profit is positive, then the TechnologyAgentwill decide to invest and join the simulation atthat time step.

Agents will be removed if they are consistentlyunprofitable for a specified length of time. Theywill also be removed if the lifetime of their tech-nology has run out.

Carbon Dioxide Tax

The CO2 tax is directly related to the fossilCO2 output for each of a technology’s Opera-tionalConfigurations. This CO2 output refers to

local emissions, not the ones indicated on thebasis of its GWP, calculated from the LCA. Forthe simulations run, a parameter sweep was per-formed using different values of the CO2 tax in or-der to find transition points where system changescould be observed.

Agent Operational Decision Making

Many of the agents have multiple Operational-Configurations. An individual OperationalConfig-uration is a list of the types and amounts of inputand output flows that pass through the agent. Forexample, a coal-fired electricity plant will use aspecified amount of coal and limestone as inputsand will generate a specified amount of electricity,ash, and several types of emissions. The sizes ofthe flows are fixed for the agents and are represen-tative of the amount of the flows in a year. Each ofthese configurations represents the use of a differ-ent feedstock. For example, if an agent wants tocofire soybean oil instead of palm oil, it will haveto switch to a different OperationalConfiguration.Furthermore, the agent can only choose a singleOperationalConfiguration per simulation time stepand cannot perform actions such as adjusting theratio of cofiring of biomass to fossil fuels.

Each agent keeps a record of profit and LCAscores associated with each of its Operational-Configurations. When the agent is initialized, itfirst collects these data by iterating through allof its OperationalConfigurations on consecutivetime steps. Essentially, the agent wakes up andassesses the economic and environmental stateof the world from its own point of view. Oncethe agent has completed this stage, it will thenselect future OperationalConfigurations on the ba-sis of the decision behavior that has been specifiedfor it. In other words, once it has looked throughits options, it will now, at every consecutive timestep, make a decision about what it thinks is best.Every time it makes a decision, it will then up-date the profit and calculate GWP for the Oper-ationalConfiguration just chosen. For every Oper-ationalConfiguration, there is only a single valuestored for profit and a single value for the calcu-lated GWP. Older values are overwritten, and theagents do not employ prediction or optimizationtechniques but just look at the last values encoun-tered. For this simulation, TechnologyAgents are



Figure 8 Global warming potential(GWP) over time; the figure showsresults for a single type of agentunder different decision types.

able to pick feedstocks on the basis of those re-sulting in the most profit or the greatest reductionin CO2-equivalent emissions. In this case, maxi-mizing profit simply means picking the cheapestfeedstock, and minimizing GWP means pickingthe feedstock with the lowest calculated LCAemissions.

In the current setup, all agents were speci-fied to have the same type of decision makingthroughout a simulation. That is, we ran a sim-ulation in which they would all try to maximizetheir profit or in which they would all try to mini-mize their greenhouse gas emissions, as measuredthrough the GWP calculated from the LCA.

Simulation Results

The simulation setup allowed two primarymeans for the system to change. First, the over-all portfolio of electricity-producing technolo-gies could change on the basis of the CO2 taxthat was chosen. All agents would have to pay atax based on their fossil CO2 emissions. Second,agents could choose particular feedstocks to max-

imize profits or environmental benefit. As for dataquality, the main focus in the case study was ongathering LCI data. The economic data gatheredwere more uncertain in comparison.

Effect of Decision MakingFigure 8 shows the effect that different styles

of decision making can have given a fixed CO2

tax. When the agents start out, they must evalu-ate the different feedstocks available to them andrecord information related to the profitability andLCA scores of each. The agent does not havea long-term memory of these values but ratherstores the last values encountered. For example,the number stored for profitability is only basedon the agent’s profit calculated in the previoussimulation time step. No prediction or optimiza-tion is done on the basis of previous values. Asshown in the graph, once the evaluation stageis complete where the agent iterates through allfeedstocks, the agents will consistently choose afeedstock on the basis of whether they are set tomaximize profit or minimize their LCA score pereach simulation time step.



Figure 9 Exploring simulation outcomes under different combinations of carbon dioxide (CO2) tax anddecision-making types. Each histogram represents 100 model runs after 20 simulated years. LCA = life cycleassessment.

The histograms in figure 9 show how accumu-lated individual agent behavior affects the overallsystem behavior. These six graphs show permu-tations of three levels of CO2 taxation and thetwo types of decision making. Each graph repre-sents multiple simulations where the agents areonly allowed to employ a single specified typeof decision making. That is, they will all, duringthe entire simulation, try to maximize their profitor minimize their LCA score, as specified by themodeler at the start of the simulation. The CO2

tax increases as one moves from the left columnto the right column of figure 9. The top row rep-resents agents trying to maximize profit, whereasin the bottom row all the agents try to minimizetheir LCA scores. For each permutation of a CO2

tax and decision-making type, a set of 100 simu-lations is completed wherein the order in whichagents act during each simulation time step is ran-domly varied. This is our attempt to see whethertrends are general and consistent, not an artifactof model construction. The simulation setup thusreflects that agents are not guaranteed to havethe same or the best trading partners. For exam-

ple, someone could have bought up all the supplyfrom a preferred partner. As described above, theagent addition process is also randomized, whichadds to the inherent variability to be expected inthe simulation results.

At the end of each simulation, we tabulatedand processed the total emissions for all theagents to calculate the ratio of fossil CO2 tototal CO2 (i.e., fossil + biogenic CO2). Thisgives a rough indicator of the amount of bio-electrity being produced versus electricity beingproduced from fossil fuels. We then placed eachof these measurements in a histogram to see thedistribution of possible system outcomes over theset of simulations for a combination of a typeof decision making and a specific CO2 tax rate.These can be understood as showing system at-tractors that indicate the range of values thatmay be encountered, rather than showing specificoutcomes.

Starting from the left column of figure 9, wesee that the agents’ decision making does notmake a great deal of difference. There is some shiftaway from complete use of fossil fuels, although



it is not necessarily significant. This is due to thedominance of large electricity producers, whichcan only cofire a limited percentage of biomass.Although they try to minimize their LCA score,the current technology as defined is incapable ofachieving significant emissions reductions. As wemove from left to right and look at the scenarioswith higher CO2 tax rates, the types of emissionsundergo a noticeable shift, indicating that morebioelectricity is being used, instead of electric-ity produced from fossil fuels. It also should benoted that the distribution of outcomes is flatterfor higher tax rates, which indicates less certaintyof one specific outcome. This is due to transitionsoccurring in the types of technologies in the sim-ulation as fossil-based electricity producers canno longer afford to pay the CO2 tax. Due to thenature of the simulation, there is no one singleway this transition could happen, which leads toa larger variety of potential outcomes over thesimulated time frame.

Essentially, we see that the ability of the sys-tem to change is defined by the dynamics of com-petition between the different technologies. Inthe simulation, new agents are added on the basisof the ratio of supply to demand and the possibil-ity that they can operate profitably. This meansthat some of the larger fossil-based electricity pro-ducers can block out the entry for some of thesmaller biobased electricity producers if the mar-ket is saturated. Some of these smaller biobasedelectricity producers can fill in the gaps as oldtechnologies go offline, although they may notbe able to make up a large percentage of totalelectricity production.

From the simulation runs, it was found that themost effective way to reduce CO2 emissions wasto impose a high CO2 tax rate, rather than onlyhaving the agents pick the feedstock that wouldlead to the lowest emissions. This means that fordrastic CO2 reductions to occur, the cofiring lim-itation mentioned above is a barrier unless it canbe overcome. Otherwise, given the assumptionsand framework used for this case study, one mustchange the portfolio of technologies to achievethese reductions.

In viewing these results, readers should notethat these outcomes were case study specific andthus tied to the definitions of technologies used.One characteristic to observe is that the tech-

nologies currently defined mostly formed linearsupply chains. Electricity-producing agents had agreater choice of possible feedstocks than agentsfurther up the supply chain. In other words, thenumber of possible supply chain network config-urations was somewhat limited. This hints thatother case studies involving different industrieswith more variety of possible connections couldlead to much different results, as a greater varietyof networks could emerge. In these cases, sup-plying information on environmental impacts toagents may have a much greater impact, as theagents would not be so constrained in their op-tions.

Readers should also remember that an agent-based model can generate a large number of data.For instance, while the model is running, we usedata logging to record all the monetary, environ-mental, and mass flows between the agents overtime. Although a subset of these data is used tocalculate a simplified LCA, many other ways ofinterpreting the data are possible. The modelercan decide to aggregate the data into system-levelindicators or to filter them to examine propertiesof individual agents. This ability to process datain different ways can aid in the evaluation of dif-ferent policies or decision-making types.

Discussion

Developers of ABM and LCA may find thismerger of tools interesting from their own per-spectives. From the point of view of LCA, oursis a modeling structure that allows for the sim-ulation of dynamic, evolving supply chains. Al-though this is not a true dynamic LCA, it doesprovide a step forward. For those in the ABMcommunity, this article shows a means to repre-sent environmental impact information from theoutside world through the inclusion of an LCIdatabase. Additionally, LCA allows for examina-tion of the supply chains that emerge from thesimulation.

What Does Life Cycle Assessment GainFrom Agent-based Modeling?

Platform for Analysis of Dynamic SystemsTraditional LCA includes several simplifica-

tions, such as linear definitions of technologies,analysis based on steady-state situations, and no



spatial differentiation of environmental inter-ventions or impacts. Heijungs and Suh (2002)talk about challenges researchers encounter inovercoming these simplifications by creating fur-ther extensions to LCA in the form of nonlinearmodels, spatially differentiated models, and dy-namic models. Although the method presentedin this article does not solve all of the issues Hei-jungs and Suh mention, it does represent a firststep in providing a platform that could conceiv-ably be extended to address them.

Nonlinear models of technologies can be usedwithin an agent-based model if appropriate rulesare defined. Such rules could state why indi-vidual agents change the production levels oftheir technologies. Additional rules could specifywhy agents decide to enter and leave the marketand invest in new technology. To define theserules, one must formalize them and program themwithin the model.

Researchers can deal with spatial differentia-tion of environmental interventions and impactsto an extent by assigning a location to each in-stance of a technology. For example, power plantscan be given specific geographical coordinates,which allows for maps of emissions to be cre-ated. This solution does not deal with diffusionof emissions but rather allows the researcher to atleast have data on the sources of emissions. Us-ing ABM with geographic information systems isbecoming more common (Gimblett 2002), andthis combination of tools could facilitate a moredynamic look at the effects of emissions on a localscale.

The analysis described in this article is dy-namic in the sense that the structure of thesystem generating electricity changes with time.Calculations are done that involve the networkstate at every simulation time step. This is not atrue dynamic LCA, according to the definitionof Heijungs and Suh (2002, 194), who stipulatethat “in making LCA a dynamic model, processesmust be specified according to the time at whichthey are active for the product under review.”

The calculations performed do not incorpo-rate time delays and thus do not track individualgoods over time but rather examine the path-ways through which goods flow. The agent-basedmodel could conceivably be extended to includewaiting times; however, one must also encode in

the model the different factors that create thesedelays. This can potentially lead to enormous datarequirements.

As shown, we have created a data represen-tation and modeling structure that allows for thesimulation of dynamic, evolving supply chains.Although it is not a true dynamic LCA, it doesprovide a step forward, and further extensionsto the modeling framework offer a platform foraddressing several of the simplifications of tradi-tional LCA discussed above.

Uncertainty AnalysisThe use of a simplified LCA within ABM has

interesting implications for a type of uncertaintyanalysis. Some LCA programs include the abilityto perform an uncertainty analysis, for instance,by specifying a probability distribution for theemissions from a process. Such a feature can beuseful, given the uncertainties due to measure-ment accuracy and system variability that are of-ten encountered in the collection of inventorydata. The uncertainty analysis can be quite im-portant for balanced interpretation of results, asit provides the researcher a means to see howuncertainty from individual system componentscan propagate to give a range of LCA scores. Forresearchers comparing two different systems, thisanalysis will indicate whether their results arestatistically distinct or whether there is so muchoverlap that they cannot confidently determinewhich one is better.

This type of uncertainty analysis only looks atone piece of the puzzle, however, as it assumesa static supply chain with predefined connec-tions between technologies. If one wants to ex-amine other possible supply chain configurations,one must manually construct them. For the im-plementation described in this article, the sup-ply chains assemble themselves on the basis ofeconomics and other system conditions. Thisflexibility can show us the range of possible sup-ply chain configurations that can exist given theconditions in the simulation. In other words, wecan test which supply chain configurations arelikely to emerge given specified simulation pa-rameters. These parameters may take the form ofa decision-making type or a certain tax regime.One may find that there is only a single possibleconfiguration or perhaps many. Additionally, one



could test whether a policy inadvertently encour-ages the emergence of undesirable supply chains.

Running the simulation multiple times withfixed parameters can lead to a type of uncertaintyanalysis in which we record the results of LCAcalculations by the agents. These results will notnecessarily be the same over time, and we canthen use these data to create a distribution ofpossible values. This distribution will show therange of emissions possible given changes in theevolving supply chain.

Because ABM is inherently a dynamic tool,this development leads to interesting implica-tions. Not only can we perform an uncertaintyanalysis based on structural dynamics, but we cando it under different system conditions, such aspolicy regimes or economic circumstances, thatwe choose to program into the model.

LimitationsThe use of an LCA within an ABM only

makes sense in certain circumstances, on the ba-sis of the type of research question being inves-tigated. This technique is especially suited forlooking at environmental performance of evolv-ing systems. For example, if one seeks to findout the environmental performance of currentsystems in a specific configuration, then a tra-ditional LCA would be more useful, due to theease of implementation and the nature of theinvestigation.

This methodology represents a level of com-plexity above a normal, traditional LCA due toadditional data requirements. The integration ofLCA into ABM can be very valuable when thereare many possible permutations of supply chainelements that can lead to nonlinear effects. Thesame is true if decision making can lead to a widevariety of outcomes. If these conditions do notexist, then this type of analysis may not be veryvaluable. This is similar to LCA itself, for whichstudies may involve different levels of detail de-pending on the needs of the research.

The current methodology employed does notattempt to find the globally optimal solution butmerely seeks a solution in which each agent is se-lecting a feedstock with the lowest cost or GWP.What we have shown does, however, provide aplatform that can be expanded to test different

theories about how system-level emissions reduc-tions may be realized.

The implementation described in this articlehas many simplifications that can be dealt with infuture work. For instance, agents only make deci-sions on the basis of current information and donot consider historical information or make anytype of prediction. There are no lead times fortechnologies, and the construction of new tech-nologies is instantaneous. Additionally, innova-tion is currently not considered.

What Does Agent-based Modeling GainFrom Life Cycle Assessment?

Developers of agent-based models can bene-fit from concepts of LCA in several ways. Twosuch benefits are explored below. First, the LCAmethodology allows for a type of structural anal-ysis to be performed. Second, the use of the LCIdatabase in this article demonstrates how thisinformation can be used to abstract the worldoutside the simulation.

Structural AnalysisThe value of ABM lies in its ability to gen-

erate complex emergent behavior. Although thisis a benefit compared to other types of modelingtechniques, it can also present a challenge, as oneneeds a means to analyze the systems that emerge.LCA is one such method. It important to realizethat LCA is really a particular implementationof a class of algorithms meant to analyze networkstructure. LCA builds on the achievements of thefields of ecology and economics, which have longbeen concerned with studying money, material,and energy flows within systems (Suh 2004). Toput this in perspective, we should not just thinkabout combining ABM and LCA. Instead, weshould think about combining ABM with var-ious forms of material and energy flow analysistechniques.

Abstracting the WorldAs already mentioned, the WorldMarket of the

agent-based model is actually a portal to an LCIdatabase. This integration means that we cansimulate a system in which we have a “lens” ofcomplex dynamic behavior where agents inter-act with a static definition of the world. One



can further refine this concept to create a moresophisticated representation of the world. Forinstance, one could allow the WorldMarket touse macroeconomic input−output tables. Sucha combination would allow the agents to operatewithin a more realistic macroeconomic environ-ment while also allowing one to simulate microe-conomic effects (Peters and Brassel 2000).

Outlook

The work shown is the result of trends inthe development of information technologies,which are changing the ways we handle massiveamounts of complex data. This is coupled withadvances in the understanding of complexity sci-ence and a shift toward what some would call thegenerative sciences, in which the question “Canyou explain it?” is replaced by the question “Canyou grow it?”(Epstein 1999). Claims of under-standing are backed by a model whose structureemerges as a result of rule sets based on thoseclaims. The model can then act as a learning toolthat allows us to test different theories about howthe system operates. This can turn into an edu-cational feedback loop, where the results of themodel inform the designer, who then updates themodel to match observed real-world behavior.

We can get closer to this vision through ex-panding the work mentioned in several ways.First, we can use more realistic decision makingand scenarios to better model the dynamics ofthe systems we are interested in. Second, the in-tegration of existing industrial ecology tools canhelp us understand the patterns that emerge andprovide the use of real-world data to define theworld that agents exist within.

More Realistic Decision Makingand Scenarios

The decision making employed by the agentsin this study was rather basic, although it couldbe made much more sophisticated. For instance,instead of only examining information from theprevious simulation tick, the agents could have amemory of information encountered throughoutthe simulation for use in their decision making.One could also use different optimization tech-niques to try to better balance complex trade-offs.

As for more complex scenarios, one could con-sider analyzing effects from economies of scale orincorporating learning effects. The learning ef-fects could take the form of reduced investmentcosts and greater efficiency with each additionalinstance of a particular technology. One couldalso test different types of policy regimes and eco-nomic circumstances to elucidate their effect onthe performance of the system. For instance, onemight investigate what could happen if the priceof a certain feedstock suddenly went up or if fi-nancial incentives were given for certain types oftechnologies.

Integrating Tools

As more processing power becomes available,it is easier to overcome limitations and criticismsof the current analysis tools through more so-phisticated calculations that are able to handleincreasing amounts of data. This is seen withimplementations such as hybrid LCA, in whicheconomic input−output tables are included as ameans to overcome the problem of limited systemboundaries.

Although it is clear that the capabilities andsophistication of tools such as LCA are expand-ing, we should also consider the shape of thisexpansion. In essence, are we making this tooldeeper or broader? Are we diving deeper and try-ing to calculate more accurate numbers, whenperhaps a different, broader insight into the sys-tem is necessary? Are we closer to understandingthe complex relationship among human equity,economy, and the environment? It is understand-able that LCA was initially developed at a timewhen only static situations could be analyzed;however, the growth in computing power givesus the opportunity to do so much more.

Figure 10 illustrates an environmental analy-sis “toolbox” commonly used for exploring prob-lems at different resolutions of flows and in-dustrial networks. Material flow analysis (MFA)examines flows of resources of a region onan aggregated mass basis. An MFA examiningflows at a substance level (e.g., chlorine, cad-mium) is referred to as a substance flow analy-sis (SFA). A process-level MFA (PMFA) is anMFA done at the level of individual processes.Physical input−output tables (PIOTs) are similar



Figure 10 Approaches of quantitative materials and energy flow analysis in industrial network systems (Suh2004, used with permission). SFA = substance flow analysis; IOA = input−output analysis; LCA = life cycleassessment; MFA = material flow analysis; PIOT = physical input−output analysis; PMFA = physical materialflow analysis.

to the input−output tables (IOTs) that describemonetary flows between industries, except thatthey document physical flows. Environmentalinput−output analysis is also similar in that itlinks environmental statistics to the informationcontained in IOTs (Suh 2004).

Although this article focuses on LCA, we be-lieve that the integration of these other industrialecology tools with the ABM platform should beexplored as well. These tools can be used to an-alyze what has emerged or to add informationabout the outside world that the agents operatewithin. For example, just as an LCI database wasconnected to the WorldMarket agent, the IOT-based tools may be connected to it as well. Al-though this has not been attempted yet, the im-plementation may have parallels to the hybridanalysis techniques that seek to combine LCAand input−output analysis, as discussed by Suhand Huppes (2005).

It is interesting that input−output analysis al-lows the creation of potentially large, evolvingnetworks containing connections based on mon-etary, mass, and environmental flows. We cangrow these networks, and, because an ontologyis used for their representation, we can under-stand what each of the edges in such a networkrepresents. We can then create algorithms thatnavigate these networks to extract information,such as an MFA. In other words, we have the

potential to generate large amounts of data aboutevolving systems that can be analyzed in multi-ple ways while coupled to real-world data. Thisvision is not a reality yet, but the work shownindicates that it may be possible.

Conclusions

This article has demonstrated a method forthe creation of a data representation and mod-eling structure that allows for the simulation ofdynamic, evolving supply chains. An LCA hasbeen used to evaluate this network in the formof a repeated accounting-type LCA. The net-works that emerge are subject to and caused bydifferent types of decision making, technologi-cal constraints, and economic factors. Some ofthe challenges addressed involve connecting thedata structures of these two tools and reducingthe computational time needed by the LCA.Some of the simplifications of traditional LCAcan be overcome either by the demonstrated workor by extensions that have been proposed.

Future development can proceed in sev-eral directions. First, the current limitedimplementation of LCA could be expanded to-ward a full-featured LCA involving improve-ments such as inclusion of more environmen-tal impacts and impact categories. Integration ofthe ABM platform with other industrial ecology



tools should be investigated as well. Work is alsoneeded on incorporating more realism into thesystems modeled. A large opportunity exists toincorporate more domain-specific knowledge inareas such as decision making and economics.For instance, one may program the agents to em-ploy multicriteria decision making or to evaluatefuture investments based on standard economiccalculations. Many further improvements can bemade, and it is hoped that the presentation ofthe work shown can lead to further discussionsof possibilities that can be realized by new typesof tool development for industrial ecology.

Acknowledgements

We acknowledge Dr. Ester Van der Voet forher supervision during the thesis on which this ar-ticle is based. Additionally, we thank Dr. ReinoutHeijungs for discussions on life cycle assessmentthat led us to overcome technical barriers andenabled this work.

References

Allenby, B.R. 1999. Industrial ecology: Policy frameworkand implementation. Englewood Cliffs, NJ: Pren-tice Hall.

Axtell, R. 2000. Why agents? On the varied motiva-tions for agent computing in the social sciences.Tech. rep. working paper no. 17. Washington,DC: Brookings Institution.

Ball, P. 1999. The self-made tapestry: Pattern formationin nature. Oxford, UK: Oxford University Press.

Bijker, W. E., T. P. Hughes, and T. J. Pinch. 1987.The social construction of technological systems: Newdirections in the sociology and history of technology.Cambridge, MA: MIT Press.

Davis, C. 2007. Integration of life cycle analysis withinagent based modeling using a case study onbio-electricity. Master’s thesis, Delft Universityof Technology, Delft, the Netherlands. http://www.industrialecology.nl/GraduationReports/2007/Davis%20MScThesis.pdf.

Ehrenfeld, J. R. 2007. Would industrial ecology existwithout sustainability in the background? Journalof Industrial Ecology 11(1): 73–84.

Epstein, J. M. 1999. Agent-based computational mod-els and generative social science. Complexity 4(5):41–60.

Gimblett, H. R., ed. 2002. Integrating geographic infor-mation systems and agent-based modeling techniquesfor simulation social and ecological processes. NewYork: Oxford University Press.

Guinee, J., ed. 2002. Handbook on life cycle assessment.Eco-efficiency in industry and science, vol. 7. NewYork: Springer.

Heijungs, R. and S. Suh. 2002. The computational struc-ture of life cycle assessment. Eco-efficiency in indus-try and science, vol. 11. Dordrecht, the Nether-lands: Kluwer Academic.

Kauffman, S. A. 2002. Investigations. New York: OxfordUniversity Press.

Nikolic, I., P. J. Beers, and G. P. J. Dijkema. 2007. Fa-cilitating interdisciplinary modelling of complexproblems. Paper presented at HICCS-40 HawaiiInternational Conference on Systems Science, 3–6 January 2007, Manoa, HI.

Nikolic, I., E. J. L. Chappin, C. Davis, and G. P. J.Dijkema. 2008. On the development of agent-based models for infrastructure evolution. Paperpresented at the International Conference on In-frastructure Systems—Building Networks for aBrighter Future, 10–12 November, Rotterdam,the Netherlands.

Peters, G. P. 2006. Efficient algorithms for life cycle as-sessment, input-output analysis, and Monte-Carloanalysis. International Journal of Life Cycle Assess-ment 12(6): 373–380.

Peters, I. and K. H. Brassel. 2000. Integratingcomputable general equilibrium models andmulti-agent systems—why and how. Paper pre-sented at the 2000 AI Simulation and Planning inHigh Autonomy Systems conference, 6–8 March,27–35. www.acims.arizona.edu/CONFERENCES/ais2000/Review/peters_i.pdf. Accessed February2009.

Projectgroep Duurzame productie van biomassa. 2006.Criteria voor duurzame biomassa productie[Final report of the project group Duurzameproductie van biomassa]. www.senternovem.nl/mmfiles/w690_tcm24-280290.pdf. AccessedFebruary 2009.

Shalizi, C. R. 2006. Methods and techniques of com-plex systems science: An overview. In Com-plex systems science in biomedicine, edited byT. S. Deisboeck and J. Y. Kresh. New York:Springer.

Suh, S. 2004. Materials and energy flows in industry andecosystem networks. Ph.D. thesis, Leiden Univer-sity, Leiden, the Netherlands.

Suh, S. and G. Huppes. 2005. Techniques for life cycleinventory of a product. Journal of Cleaner Produc-tion 13(7): 687–697.



Van der Voet, E., L. van Oers, C. Davis, R. Nelis, B.Cok, R. Heijungs, E. Chappin, and J. B. Guinee.2008. Greenhouse gas calculator for electricityand heat from biomass. CML report 179. Leiden,the Netherlands: Institute of EnvironmentalSciences (CML), Leiden University. http://www.leidenuniv.nl/cml/ssp/publications/co2_tool.pdf.Accessed February 2009.

Waldorp, M. 1992. Complexity: The emerging science atthe edge of order and chaos. NY: Simon and Schus-ter, USA.

About the Authors

Chris Davis, Igor Nikolic, and GerardP. J. Dijkema are affiliated with the Fac-ulty of Technology, Policy and Manage-ment within Delft University of Technol-ogy in Delft, the Netherlands. Chris Davisand Igor Nikolic are Ph.D. researchers,and Gerard P. J. Dijkema is an associateprofessor.


Documents

Integration of Life Cycle Assessment Into Agent-Based Modelingre.indiaenvironmentportal.org.in/files/Integration of Life Cycle.pdf · Integration of Life Cycle Assessment Into Agent-Based