An Agent-Based Computational Economics Approach to Forecasting Freight Flows for the Chicago Region J. Gliebe, J. Doyle, K. Wies, C. Smith, K. Shabani, M. Outwater TRB Innovations in Travel Modeling -- Baltimore April 29, 2014
Acknowledgments David Wu, Lehigh University Sunil Chopra, Northwestern University Matt Roorda, University of Toronto Craig Heither, CMAP Jeff Keller, 2
Overview Motivation, questions, objectives Theoretical underpinnings Linkages between industries Model structure and elements Procurement market game Numerical results What s next 3
Motivation Tool needed to address economic development questions on freight policy and investments Limitations of commodity flow data Future looks much like today Inability to respond to economic stimuli Fuel prices Productivity improvements Trade policies, tariffs, etc. Insensitive to changes in capacity investments or behavioral policies 4
Questions a traditional freight forecast method can t answer How will the freight industry s business model change if: Chicago s CREATE program to address a national rail bottleneck is completed? A new multi-modal terminal is built in Iowa? The deep-water port in Prince Rupert, BC is expanded? U.S. trade policies with China are changed? 5
Modeling performance objectives High-level scenario analysis of assumptions affecting the formation of supply chains in a global economy Endogenous modeling of origin-destination flows Theoretically plausible shifts in the magnitude and direction of flows and modal usage in response to both regional and extra-regional stimuli Circumstances (scenario assumptions) under which various outcomes are likely to occur Predict emergent patterns of response that the analyst might not expect, but should consider 6
Theoretical drivers Individuals make supply chain decisions Imperfect information, culture/custom, affinity, imitation, satisficing Everyone wants to make/save money, but mathematical optimization is rare Variation in value systems across markets Cost savings, time responsiveness, need for centralized control Variation in market mechanisms Ad hoc bi-lateral agreements, auctions/bidding, collusion, oligopolies 7
Agent based computational economics (Tesfatsion et al 2006) Agents interact in simulated world following simple rules, often cooperation games Repeated encounters lead to learning, adaptive strategies Emergent behavior Used to study many adaptive, complex systems Source: http://computationallegalstudies.com/2010/07/27/agents-of-changeagent-based-models-and-methods-the-economist/ 8
9
Economic drivers 10
Supply chain networks: linkages between industries, firms, individuals (U.S. BEA input-output accounts) Consumer centric view Producer centric view 11
1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 Multi-factor productivity (U.S. BEA Index) I n d e x V a l u e s 450 400 350 300 250 200 150 100 50 0 NAICS 3341 - Computer and peripheral equipment 2002 = '100' Multifactor productivity Labor productivity Capital productivity Intermediate purchases productivity Output 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 I n d e x V a l u e s 140 120 100 80 60 40 20 0 NAICS 3117 - Seafood product preparation and packaging 2002 = '100' Multifactor productivity Labor productivity Capital productivity Intermediate purchases productivity Output 12
Sourcing: the heart of the supply chain 13
Initializing agents Buyer Attributes NAICS Size (# employees) Freight zone Output commodity Input commodity Input commodity requirements ($ annual purchase) demand B S Seller Attributes NAICS Size (# employees) Freight zone Output commodity Production level ($ annual output) capacity Buyer Preferences Min. cost vs. Max. response Risk spreading Centralized distribution Vertical Integration Buyer-Seller Matching Game Seller Cost-Service Bundle Shipment sizes Average shipping times Distribution centers Mode Cost 14
Procurement market game Inspired by trade network game (McFadzean et al, 2001) Iterative pairwise encounters (solicitation bid) B Decision Matrix Buyer Seller Decision Payoffs Decision Payoffs Yes Utility of Transaction Yes Revenue of Transaction No Expected Utility if another supplier must be chosen No Expected Revenue of holding out for other (more lucrative) buyers S Actual payoffs revealed at end of round after other potential pairwise games are played Agents rate their experiences with other agents 15
Agent payoffs, memory and learning Expected values formed dynamically through game play Cooperate This is a favored trading partner Defect Other trading partners are preferred B Payoff Matrix Seller Buyer Cooperate Defect Cooperate 1.0 0.6 Defect 0.6 0.6 Bi-lateral agreement game form (others possible) Payoff factors are applied to calculated utilities/revenues Defect payoffs discount what might have been or could be Each buyer agent remembers outcomes of pairwise games with each seller agent, and vice versa S 16
Constraints Sellers assumed to be production constrained May not be able to supply 100% of buyer s needs Downgraded commensurately by buyer Buyers may be assumed to be risk averse Max. fraction to be purchased from any one source Max. fraction to be shipped through any one port Will try to find multiple sources Implications of many buyers and few sellers May settle for non-favored sellers May relax max. fraction constraints 17
Examples of game chatter from log file Buyer 851828 buy from Seller 30649 quantity: 2316 Buyer 4427804 buy from Seller 30324 quantity: 776 Adjusting BuyerTradePayoff 883760 <- 30324 quantityrequested: 1751 ; quantitydelivered: 1685 fractionalsatisfaction: 0.962307 Suspending maxfrac for buyer 845650 30621 diminishing offer from 30347 for 21423 to 13524 30621 rejecting offer from 4427851 for 263 Buyer 4421578 asks Seller 30615 for 33 Seller 30615 OVEREXTENDED: capacity 3166894 shortfall 28331 Buyer 4417894 Tag CONSTRAINT 0 (30610) reducing supplemental offer from 341609 to 95 Buyer 4417894 asks Seller 30610 for ADDITIONAL 95 18
Example NAICS Commodity 2122A0 Iron, gold, silver, and other metal ore mining Top consumers from I-O tables NAICS 325510 Paint & Coating Manufacturing 19
Consumer locations and characteristics for iron, gold, silver, and other metal ore mining 18,762 20
Producer locations and characteristics for iron, gold, silver, and other metal ore mining 269 21
Comparison of Chicago area producer and consumer locations for iron, gold, silver, and other metal ore mining (Lots of imports) 22
Consumer locations and consumption of iron, gold, silver, and other metal ore mining 23
Producer locations and production of iron, gold, silver, and other metal ore mining 24
All attempted trades and final trades with one seller Fontana Metal Works 25
All attempted trades and final trades for one Chicago buyer Cicero Paints & Coatings 26
Distance distributions for all pairs, trading pairs, final iteration pairs 27
What s next Now through June 2014 Address computational issues Sampling and/or representative agents to reduce combinatorics, developing closure criteria Adding representative agents for foreign trade Complete working model July 2014 June 2015 Improve resolution on non-transport prices, tariffs Sensitivity testing and fine tuning of parameters Comparisons with commodity flow and other data Scenario testing 28
Questions and Answers
Contacts www.rsginc.com John Gliebe, PhD Senior Consultant, john.gliebe@rsginc.com Kermit Wies, PhD Deputy Executive Director for Research and Analysis, CMAP kwies@cmap.illinois.gov
Reserve slides to follow if needed for Q&A 31
Cost function Ordering Cost Transport and Handling Cost Damage Cost Inventory in-transit cost Carrying Cost Safety Stock Cost (Source: Cambridge Systematics, 2011) 32
Levels of response sensitivity in forecasting Stimulus: Change in price of an input commodity A 1. Switch suppliers for input commodity A? Part of model s basic response set-pmg 2. Switch suppliers for input commodity B? Requires budget-awareness. Would need to be sub-problem within PMG. Perhaps too complex defer to future Supply chain activity network propagation 3. Pass cost change along in price of output commodity? Change I-O coefficients? Change output levels? Change in final demand? 4. Cost is global and changes factors of production? 33