Slide 1

Design of an Agent-Based Computational Economics Approach to Forecasting Future Freight Flows for the Chicago RegionOctober 21, 2013Prepared for:TRB SHRP 2 C43 Symposium on Innovations in Freight Modeling & Data Prepared by:John Gliebe, RSG, Inc. (Presenter)Kermit Wies, CMAPColin Smith, RSG, Inc.Kaveh Shabani, RSG, Inc.Maren Outwater, RSG, Inc.OverviewBackground and motivationAgent-based computational economicsOverview of current model system Overview of proposed model design Procurement market game Next steps

2Background The Chicago Metropolitan Agency for Planning (CMAP) has recently developed a two-tiered modeling system to analyze regional freight traffic Original work: FHWA BAA Research project (2011), Cambridge Systematics, University of Illinois-Chicago, RSG, Inc.Address complexities of freight modeling: commodities produced and consumed, local pick-up and delivery, warehousing and tour formationThe first tier is referred to as the mesoscale model or national supply chain model and represents freight flows between the CMAP region and the rest of the U.S. The second tier model operates within the region, modeling local freight-hauling truck movements using a tour-based microsimulation and has been dubbed the microscale model or the tour-based truck model Commodity flows, including future freight flows, are derived from those found in the Freight Analysis Framework, Version 3 (FAF3) data products, developed by the Federal Highway Administration (FHWA).

3The purpose of this project is to develop a third major component to the modeling system, an extension of the mesoscale model that would enable CMAP to forecast future freight flows independent of FAF3, based on macroeconomic indicators as well as a host of other stimuli. This document presents the conceptual design of this mesoscale extension, which features an agent-based computational economics modeling component.

3MotivationFor policy and planning sensitivity analysis, CMAP would like a tool to systematically vary forecasts to reflect:Potential changes in macroeconomic conditions (e.g., foreign trade levels, price of crude oil);Large-scale infrastructure changes (e.g., port expansions, new intermodal terminals);Technological shifts in logistics and supply chain practices (e.g., near-sourcing, out-sourcing, productivity enhancements); and Other assumptions and scenario inputs related to the economic competitiveness of the Chicago region and its infrastructure investments.To support these objectivesFuture freight forecasts will need to be produced endogenouslyfocus on producer sourcing decisions to meet production levelsForeign trade and macroeconomic conditions will need to included in the producer sourcing decisionsPrice signalstransport and logistics costs and other costs will need to be part of sourcing decisions4Theoretical Drivers Behind ApproachSupply chain decisions are made by individuals on behalf of the companies for whom they work, characterized by:Imperfect informationCultural baggagePersonal affinity for particular business partnersLimited search efforts based on custom, imitation, satisficing behaviorPurchasing agents want to save costs and selling agents want to maximize revenues, but It may not be realistic to assume that agents will make mathematically optimal decisionsVariation in value systems across markets (examples):Emphasize cost savings for bulk commodities with low storage costsEmphasize frequency of shipments for perishable commoditiesPractice vertical integration (in-house sourcing) for complex commodities, such as components of high-precision medical equipmentVariation in market mechanismsAd hoc bi-lateral agreementsAuctions and biddingCollusions, side agreementsVarying levels of competition

What is agent-based computational economics?ACE research characterized by rigorous study of economic systems through computational experimentsBottom up approach in which individual agents are simulated in a virtual world in which they make decisions, interact with and react to each other, and patterns emerge from the collective actions of many agents

6

Interactions between agents typically take the form of cooperation gamesMethodological kinship with complex systems studies in social and natural sciencesElectric power tradingSocial choice and votingRacial segregation in housingSchool choiceHabitat destructionHoneybee swarmsSource of graphic: http://computationallegalstudies.com/2010/07/27/agents-of-change-agent-based-models-and-methods-the-economist/6ACE: Answering questions about complex systemsWhy have certain global regularities emerged and persisted despite the absence of centralized planning and control, while other global outcomes have not been observed? How are trends in supply-chain and logistics practices, such as insourcing, outsourcing, and near-sourcing influenced by privately held values and beliefs regarding various forms of uncertainty, asset specificity, and commodity attributes? How much is simply imitation?What types of micro-level dynamics of individual traders lead to the collective patterns market behavior that we observe?Which agents in the supply chain network have the greatest influence on other agents (commodities, industries)? Are there ties between agent/industries that may be important to assessing regional competitiveness and the likely trends in future freight flows?How can good economic (infrastructure) policies be designed to achieve their intended effect?Road pricing, traffic flow management, trade tariffs, port capacity expansions

7Current Mesoscale Freight Model 8Synthesizes a list of businesses in Chicago, the rest of the US, and an international sampleConnects suppliers to buyers based on the commodities produced by the supplier and consumed by the buyerDistributes commodity flows amongst the paired suppliers and buyersFor each buyer/supplier pair, selects whether shipments are direct or involve intermediate handling (intermodal, distribution center) For each buyer/supplier pair, converts an annual commodity flow to shipments by size and frequencyIdentifies the mode for each leg of the trip from supplier to buyer and the transfer locationsThe local deliveries and picks up in the Chicago area are simulated using a truck touring modelNational ScaleRegional Scale

National Scale Model Sequence9

Major Changes to Mesoscale Model DesignAdding attributes to describe firms/establishments, based on operating typologiesCommodity flow data no longer basis for predictions, but used in calibration and testing as benchmarksBusiness linkages, shipment demand, and mode determined through a joint decision framework, based on the outcome of agent-based simulations Outcome are shipments that feed directly into the truck touring model (simulation of delivery and pickup)

10Model Design Overview and Integration11

In terms of integration into the mesoscale model, the extension illustrated in Figure 3 would:Pick up at the point where firms have been synthesized by the current model;Modify and augment the attributes of each firm and create purchasing and selling agents;Replace the current mesoscale models static matching of buyers and sellers with the agent-based procurement markets approach;Integrate the choice of supplier firms with the choice of shipment modes and sizes and choose distribution channels, using the same method found in the current mesoscale model; andReturn a set of annual commodity shipments by mode-path and average shipment size.

11Synthesize FirmsStart with existing method of firm synthesisWithin the CMAP region, firms are treated as establishments in that they are situated in a single location and function as establishmentsOutside of the CMAP region, representative firms will be created to represent a single industry, and region/country (FAF zone)E.g., Wyoming Coal ProducersFirm attributesIndustry Code (NAICS)CPB Zone (County Business Patterns zone used during supplier selection)FAF Zone (country/region)CMAP modeling zoneCommodity Type(s) Produced (SCTG)Size (number of employees)Production capacity (commodity units produced per year)

12Assign Firm Types and PreferencesPurpose to define firms sourcing preferences (tradeoffs) for various combinations of source offerings (attribute bundles)Commodity Service Offerings Unit cost / total costAverage shipment timeFrequency of shipments / Average shipment sizeProximity of supplierPerceived reliability of the supplierPerceived quality of the suppliers commodity (assert for certain scenarios)Firm Operational TypesEfficiency vs. Responsiveness: Is commodity innovative or functional?Geographic Proximity: Are there preferences for near-sourcing vs. far-sourcing?Centralization Tendencies: Is commodity likely to utilize warehousing and distribution systems?Vertical Integration Tendencies: Is commodity likely to be produced in-house?13Assign Firm Commodity Production LevelsU. S. Bureau of Economic Analysis (BEA) data will be used to estimate the total dollar-value of output commodities based on firm sizeAccount for production cost differences for non-U.S. countriesFor imports, BEA reports producer prices at U.S. port of entry in U.S. dollars14CommodityIndustryTotal IntermediateTotal Final UsesTotal Commodity OutputIndustry AIndustry BIndustry CCommodity A5010602585Commodity B406020120-10110Commodity C-10110120100220Total Intermediate9070140300--Total Value Added402040-115-Total Industry Output13090180--400Calculate Input Procurement RequirementsBEA Input-Output (I-O) tablesUse tables after redefinitions to represent only direct inputsNormalized becomes Direct Requirements table

15

(partial table shown)Recipe read down column corresponding to each producer industryTotal Commodity Output: Rightmost column in the Use table

Adjusted tables refers to Make and Use tables after redefinitions and reallocation of commodities to the primary industries that produce them. When this is done, the Use table only include direct inputs, not indirect inputs (the inputs of other inputs). This is what we want and is better for modeling freight flows because it will help us to maintain a truer depiction of freight flows.

Direct Requirements tables are derived from the Use tables and represent the proportion of a dollar of output in industry j (column) that is provided by commodity i (row). This is what we want to use.

Total Requirements tables include intermediate inputs, i.e., both direct and indirect requirements and come in three forms: commodity by commodity, industry by commodity, and industry by industry.

Imports table is supplemental and imputed by BEA. It is available in the redefined table form. This could be used to derive calibration target values for imported vs. domestic sourcing.

15Estimate Transport and Non-Transport CostsTransport and Logistics CostsUse skims from the multi-modal network model and unit costs created as part of the current mesoscale model to provide transport and logistics costs, composed of:Ordering costTransport and handling costDamage costInventory in-transit costCarrying costSafety stock cost

Total CostsUse FAF3 estimates of total shipment values between FAF zones to provide a total cost figure

Non-Transport/Logistics Costs = Total Costs Transport Costs

16Match Buyers and Sellers, Bundling Service AttributesCreate buyer agents with preferences for bundled cost-service attributesFor each of the 43 commodity types under consideration, a procurement market model will be runThis is done through a procurement market gameThe objective of this step is to find suppliers for every commodity input required by buyersOutcome is joint choice of supplier, shipment size, distribution center use and mode-path17

Procurement Market Game (PMG)Research literature in supply chain sourcing decisions focuses on auction mechanisms that can be used to optimize outcomesE-procurement systems require efficient, robust algorithms (algorithmic game theory)Common objectives are to induce suppliers to bid at true cost, avoid collusion, and other forms of strategic lyingExample: 2nd Price Sealed Bid (Vickery 1961)Appropriateness of auction mechanisms for Mesoscale Freight ModelIndustry and commodity-specificNot necessarily applicable to smaller and less technologically advanced firmsTypically designed for optimizationWont necessarily capture idiosyncratic behavior of agent preferences, habits and beliefsACE approaches offer a more general, flexible frameworkPMG inspired by Trade Network Game (TNG) (Tesfatsion, McFadzean, Iowa State U.)Agents are buyers, sellers and dealers (buy or sell)2 x 2 Payoff matrix cooperate or defect labeling (e.g., Prisoners Dilemma) Evolutionary programming framework (genetic algorithm)Multiple rounds of pairwise tradesAgent expectations about other trading partners are updated after each round based on outcomes of all pairwise trading gamesMarket properties emerge through iterative play

18Initializing Agents19

B

SBuyer AttributesNAICSSize (# employees)FAF ZoneOutput commodityInput commodityInput commodity requirements ($ annual purchase) demandSeller AttributesNAICSSize (# employees)FAF ZoneOutput commodityProduction level ($ annual output) capacityBuyer PreferencesEfficient vs. ResponsiveNear-source vs. Far-sourceCentralized DistributionVertical IntegrationSeller Cost-Service BundleShipment sizesAverage shipping timesDistribution centersModeCost

Shipment sizes, average shipping times, usage of distribution centers19Round 120

B

S

B

B

B

B

S

SRound 221

B

S

B

B

B

B

S

SRound 322

B

S

B

B

B

B

S

SExample PMG: Trade Scenario ALarge buyer L and a small buyer S who are both in the packaged foods industry, commodity code, CC=1Each buyer needs to purchase a quantity of an input commodity, seafood, commodity code, CC=2. Both buyers are in the geographic zone, GZ=1Three sellers: foreign (F), domestic (D), and local (L), each offering different bundles of unit costs and shipping time23

Buyer L represents a large manufacturer who is of the type that values timely delivery and is willing to pay a little more for it. Buyer L being smaller needs to be more cognizant of costs, even though timeliness is important for obvious reasons of freshness. The weights that each buyer places on the (dis)utility of cost and shipping time are shown in Table 4.1. In this example, an opportunity cost of -7.0 is the payoff at the end of a trading round if a buyer fails to secure a seller. Also, in this game it is assumed that there can only be one seller for each buyer, so at least one seller will be left out of the market.

23Pairwise Trade L-FLarge Buyer and Foreign Seller24

Buyer L would reject the trade based on the expectation of getting a better deal (more positive utility) from one of the other two sellers. Seller F would say yes to the deal based on the expected revenue being greater than what could be expected from the other buyer, taking into account the competition with two other sellers. 24Pairwise Trade S-DSmall Buyer and Domestic Seller25

It is important to note here that it is assumed that pairwise trades take place in no particular order. Within a round, the agents in each pairwise game have no knowledge of any other pairwise game. Table 6 below shows the outcomes of a pairwise game between the small buyer and the domestic seller. In this case, the buyer says yes to the deal, but the seller say declines, holding out for a higher expected revenue from the large buyer.

25Pairwise Trade L-LLarge Buyer and Local Seller26

In this game, both parties agree to trade and receiving the commensurate utilities/revenues from the trade. 26Scenario A ResolutionAnd so on All pairwise combinations (2 x 3 = 6) are calculated and expected payoffs for each game are updated based on these pairwise outcomes

Only partnership formed was between Buyer L ("large") and Seller L ("local").Under an assumption of mutual exclusivity, an initially favorable L-D match was superseded by L-L (slightly better for the buyer)Buyer S ("small") was outbid after holding out for the preferred provider ("local").Buyer S was rejected by all of the sellers, who were holding out for Seller L ("large").During the second round, buyers and sellers would update their beliefs about the probability of a successful trade, which should result in a second alliance forming between Buyer S and Seller D ("domestic").Seller F ("foreign") is priced out of this market for fish, but could become competitive in a different scenario if cost structures or preferences were to change.

27

Expectations of PMGReplicating what actually goes on in a procurement market is challengingDifferent payoff matrices may be defined to capture different styles and assumptions on the bilateral trade, resulting in different emergent behaviorWe may create 3-5 general types of games to represent commodity markets of similar typesBuyers will outnumber sellers in the majority of marketsPair-feasibility criteria will be developedStopping criteria to be determined, but providing suppliers to fulfill every buyers input needs will be minimumconvergent solutions preferredSellers will have capacity constraintsSome buyers will want to spread risk and choose multiple sellersCost structure assumptions and parameters, and utility preference weights will be highly influential, thus a large part of the development timeFAF3 flows will be used for benchmarking and calibration of aggregate results

2828Planned Scenario TestingImpacts of full implementation of Chicago's CREATE programImpacts of implementation of Midwest Intermodal Hub in IowaImpacts of expansion of Port of Prince Rupert, BCImpacts of reduction/increase in U.S. Trade with China

29

Source: http://www.createprogram.org/linked_files/ProjectMap_print.pdfNext StepsDevelopment of PMG test bed (ongoing)Incremental testing of variables, algorithms and assumptionsAssess convergence properties, reasonableness of results, computational time, realism of portrayed competitionStart with single commodity procurement market; gradually try additional markets, look for commonalities and generalizabilityStart with simpler single-cost variables in payoffs, gradually add other variablesTest different utility weighting parameter valuesTest full combinatorial treatment vs. filtering and samplingExperiment with assumptions information known to agentsExperiment with other algorithmic assumptions30

Colin SmithRSGcolin.smith@rsginc.com802-295-4999John GliebeRSGjohn.gliebe@rsginc.com240-283-0633

San DiegoEvansville31Kaveh ShabaniRSGkaveh.shabani@rsginc.com801-736-4100Maren OutwaterRSGmaren.outwater@rsginc.com414-446-5402 www.rsginc.comKermit WiesCMAPkwies@cmap.illinois.gov312-386-8820Craig HeitherCMAPcheither@cmap.illinois.gov(312) 386-876831Reserve Slides for Q & AUse only if needed to aid in answering questions32Firm Synthesis33Firms are synthesized for the entire U.S. with a high level of industrial sector detail, and across several employment size categories

Data:County Business Patterns: business establishments by employment size category and industrial sector for each countyInput/output data (US BEA) to tag each establishment with the commodity that it produces/consumes (interested in the commodities that require transportation)

NAICS Groups1-1920-99100-249250-499500-9991,000-2,4992,500-4,999Over 5,000TotalAgriculture631,70383,32811,9412,8979543824613731,264Mining\Construction774,69773,60710,0903,7112,3631,9001,7441,716869,828Manufacturing228,38174,45118,9426,1702,38482814349331,348Transportation\ Wholesale\Retail1,518,135214,95634,0827,3051,53639371341,776,512Information\Finance\ Professional Services2,094,868186,14032,43110,1414,3361,737295972,330,045Education\Healthcare731,344110,50420,1204,5232,1681,748435157870,999Entertainment\ Recreation\Food558,052186,14011,0691,5225762694315757,686Other Services694,64045,3773,40954816340123744,192Total7,231,820974,503142,08436,81714,4807,2972,7892,0848,411,874Application Development in RScripting SoftwareR version 2.14RuntimeTotal run time is 80-90 minutes

34HardwareManufacturer:HP Z200 WorkstationProcessor:Intel Core i7 CPU 870 @ 2.93 GHzInstalled RAM:12.0 GBOS:Windows 7 Professional (64-bit)Model ComponentRun Time (minutes)NotesFirm synthesis13Synthesize 8 million firms and choose buyers (7.5 million) and suppliers firm types (1.4 million) for CMAP simulationSupplier selection24Match supplier firm types for about 3 million firmsSupply chain and goods demand19Apportion FAF flows for 3 million buyer supplier pairs and locate 8 million firms to mesozonesDistribution channel1.0Predict distribution channels for 3 million buyer-supplier pairs using logit sharesShipment size1.5Estimates annual shipment size and frequencyMode-Path selection20Evaluation of annual logistics and transport costs for 54 mode-pathsVehicle choice and tour pattern1.5Daily simulation for 300k deliveries\pick-ups from warehousesStop clustering and sequencing1.5Clusters and sequences stops on toursStop duration0.2Estimates stop durationTime of day1.5Constructs tours from start time and stop durationCost Function (Current Mesoscale mode, improved equation)Variable or ParameterDescription or Interpretation (of Parameters)SourceGmnqlLogistics cost (shipper m and receiver n with shipment size q and logistics chain l)Calculated in the modelQAnnual flow in tonsFAFqShipment size in tonsVariableb0qlAlternative-specific constantParameter to be estimatedb1Constant unit per orderParameter to be estimatedTTransport and intermediate handling costsnetwork skims, survey datab2Discount rateParameter to be estimatedjFraction of shipment that is lost or damagedSurvey data or assumed valuevValue of goods (per ton)FAF datab3Discount rate of goods in transitParameter to be estimatedtAverage transport time (days)Lookup table (or skims), survey datab4Storage costs per unit per yearParameter to be estimatedb5Discount rate of goods in storageParameter to be estimatedaConstant, set safety stock a fixed prob. of not running out of stockSurvey data or assumed valueLTExpected lead time (time between ordering and replenishment)Lookup table (or skims) , survey datasQStandard deviation in annual flow (variability in demand)Survey data, assumed valuesLTStandard deviation of lead timeLookup table (or skims), survey data(source: Cambridge Systematics, 2011)Inventory in-transit costCarrying CostSafety Stock CostTransport and Handling CostOrdering CostDamage Cost

Levels of Response Sensitivity in ForecastingDifferent levels of response sensitivity can be incorporated in the model design (not mutually exclusive)

361. Switch suppliers for input commodity A?2. Switch suppliers for input commodity B?3. Pass cost change along in price of output commodity?Stimulus: Change in price of an input commodity A4. Cost is global and changes factors of production?Part of models basic response set-PMGRequires budget-awareness. Would need to be sub-problem within PMG. Perhaps too complex defer to futureSupply chain activity network propagationChange I-O coefficients?Change output levels?Change in final demand?Model Design Overview and Integration37

In terms of integration into the mesoscale model, the extension illustrated in Figure 3 would:Pick up at the point where firms have been synthesized by the current model;Modify and augment the attributes of each firm and create purchasing and selling agents;Replace the current mesoscale models static matching of buyers and sellers with the agent-based procurement markets approach;Integrate the choice of supplier firms with the choice of shipment modes and sizes and choose distribution channels, using the same method found in the current mesoscale model; andReturn a set of annual commodity shipments by mode-path and average shipment size.

37Stone mining/quarrying (supplier perspective)38

Stone mining/quarrying (1st, 2nd order ties)39

Stone mining & quarryingStone mining/quarrying (1st, 2nd, 3rd order ties)40

Iron and steel ferrous alloysStone mining & quarrying