^-*^!!^ <>> — j^ l«^' iAUG ALFRED P. WORKING PAPER SLOAN SCHOOL OF MANAGEMENT EXPERIMENTAL EVIDENCE OF DETERMINISTIC CHAOS IN HUMAN DECISION MAKING BEHAVIOR John Sterman, Erik Mosekilde, Erik Larsen REVISED June, 1988 WP# 2002-88 MASSACHUSETTS INSTITUTE OF TECHNOLOGY 50 MEMORIAL DRIVE CAMBRIDGE, MASSACHUSETTS 02139 rr,-: 18 1988 EXPERIMENTAL EVIDENCE OF DETERMINISTIC CHAOS IN HUMAN DECISION MAKING BEHAVIOR John Sterman, Erik Mosekilde, Erik Larsen REVISED June, 1988 WP// 2002-88 D-3957-1 Experimental Evidence of Deterministic Chaos in Human Decision Making Behavior John D. Sterman*, Erik Mosekilde**, Erik Larsen" March 1988 Revised June 1988 Abstract An experiment with a simulated microeconomic system demonstrates that the decision- making processes of human subjects can produce deterministic chaos. a commodity production-distribution system systematically suboptimal. estimates show the model A is model of the to minimize costs. Participants managed Performance, however, was subjects' decision rule is proposed. Econometric an excellent representation of the actual decisions. Simulation of the estimated rules yields stable, periodic, quasiperiodic, and chaotic solutions. Analysis of the parameter space reveals a complex structure including mode-locking, a devil's case, and fractal basin boundaries. Implications for modeling human systems stair- are explored. Massachusetts Institute of Technology, Sloan School of Management, Cambridge, MA 02139 ** Physics Laboratory HI, Technical University of Denmark, DK-2800 Lyngby, Denmark 2 Sterman, Mosekilde, and Larsen The prevalence of deterministic chaos and other nonlinear phenomena in physical, chemical, and biological systems has prompted speculation that these dynamics may occur in Indeed, numerous models have economic, managerial, and other human systems as well (1). shown how economic systems might produce chaos But the significance of these (2). important theoretical developments hinges on whether the chaotic regimes Further, the decision rules postulated in these models have not region of parameter space. been tested experimentally. Despite intriguing series (3), it is the realistic lie in difficult to resolve efforts to identify chaos in such issues by empirical means alone often unavailable, or too short relative to the frequencies of interest. economic time Data (4). series are Measurement error and process noise in social and economic data are significant, further complicating empirical analysis (5). A complementary approach is based on laboratory experiments provide a simulated environment for the study of decision-making results of human The experiment systems give firms We report here common economic the flexibility by establishing a decentralized network of inventories demand "Beer Distribution Game," at is MIT The experiment The experiment is here, the a role-playing simulation of a typical production-distribution to introduce students of management to economic dynamics and computer simulation, the game has been widely used for nearly three decades conducted on a board which portrays production and distribution of beer (figure manipulated by the players as the game 1). Beer proceeds. wholesaler, distributor, and factory (R, is to buffer and production of commodities and for business cycles and other fluctuations in industrial economies (7). Customer demand Such Production-distribution systems have long been associated with manufactured goods. Developed context. creates a simulated industrial production-distribution system. unanticipated variations between the retailer, which models one such experiment which demonstrates that the decision-making processes of subjects can produce deterministic chaos in a system. (6). in is (8). in a simplified fashion the represented by markers which are Each brewery consists of four W, D, represented by a deck of cards. F). One subject sectors: manages each sector. Each simulated week customers order Sterman, Mosekilde, and Larsen beer from the who retailer, 3 ships the beer requested out of inventory. orders from the wholesaler, who likewise ships out of inventory. orders and receives beer from the distributor, The factory. factory brews the beer. Each week delays. (factories decide the subjects how much who The Similarly, the wholesaler and receives beer from the in turn orders At each stage there are shipping and order receiving must decide how much to order from their immediate supplier beer to brew). Subjects seek to minimize cumulative costs incurred during the weeks). retailer in turn game (36 simulated Inventory holding costs for each sector are $.50/case/week, and backlogs of unfilled orders cost $ 1 .00/case/week. Due avoiding backlogs. unfilled orders may Subjects must therefore keep their inventory low while and shipping to the order receiving build up. The may lag in receiving beer lags, a substantial supply line of vary: if the wholesaler can cover the retailer's orders, the retailer will receive the beer desired after four weeks. But if the wholesaler has run out, the retailer must wait until the wholesaler receives additional beer. Only the factory, the primary producer, faces a constant delay in acquiring inventory (for simplicity there is highly simplified, no it limit to production capacity). nevertheless shares many negativity constraints on orders and shipments: if inventory is is features with real economies, including The multiple feedbacks, nonlinearities, and time lags. but Although the experimental system nonlinearities arise from non- shipments normally equal incoming orders, depleted, shipments must equal zero, and the unfilled orders are held in the backlog for future delivery. The experiment follows standard protocols initialized in equilibrium. cases per week. demand Each inventory contains 12 The system to 8 cases per Forty-eight week trials (6, 9) is in and cases. is detailed in (10). Customer demand week is initially 5. (192 subjects) were analyzed. lower costs than the trial is disturbed by an unannounced step increase in customer Trials in accounting errors were discarded, leaving eleven (44 subjects). slightly Each full which subjects made This subset generated sample, indicating that these subjects understood and 4 4 Sterman, Mosekilde, and Larsen performed best MIT Participants in the experiment. were graduate and undergraduate students and senior executives from a number of U.S. firms. Results are summarized in table behavior is minimize regularities, suggesting subjects 1; figure 2 shows a typical More costs. used a common initial The participants' it is not surprising interesting, the results exhibit strong heuristic in ordering. and inventory are dominated by a large amplitude fluctuation. required to recover trial. Given the complexity of the system significantly suboptimal. that subjects are unable to inventory levels. 2. Amplification: orders rises steadily from customer to retailer to factory. 8 at On 1. Oscillation: Orders average 21 weeks are The amplitude and variance of Customer orders increase from 4 to cases/week; responding to this disturbance, factory orders rise from 4 to a peak averaging 32 cases. 3. Phase Orders tend to peak lag: later as one moves from retailer to factory (not surprising since orders propagate upstream through decision-making and processing delays). We next estimate a model of the subjects' decision rule (see (10) for details). We hypothesize that subjects utilize the anchoring and adjustment heuristic to determine orders (11). Anchoring and adjustment quantity is estimated by adjusting for factors person's weight first is common recalling a which may be may be a judgmental strategy known which an unknown reference point (the anchor) and then less salient or whose effects are obscure. estimated by recalling the subject's differences in height, girth, etc. in own weight and For example, a adjusting for Anchoring and adjustment has been documented in a wide variety of decision-making tasks (11, 12). Here the anchor is expected demand D^, the subject's forecast of incoming orders. Replacement of incoming orders, ceteris paribus, maintains inventory at The asymmetric cost function means nonnegative value S'. the desired stock S' line of unfilled orders. received. The supply Finally, orders O current value. subjects should maintain their inventory at a Thus an adjustment and actual stock its S. line to the Subjects SL is anchor arises from the discrepancy between may also consider a fraction P of the supply the accumulation of orders placed but not yet must be nonnegative (order cancellations are not allowed). Thus, Sterman, Mosekilde, and Larsen Ot= MAX(0, 5 D^t + a(S' Sf - where the stock adjustment parameter pSLj)) (1) a represents the fraction of any discrepancy corrected each time period. The adjustment term a(S' regulate the stock. Sf - pSLj) creates two negative feedback loops which Discrepancies between the desired and actual stock induce additional orders until inventory reaches the desired value. on order induces additional orders Likewise an insufficient quantity of goods - depending on If p = orders placed are forgotten until they arrive, causing overordering and instability. If until the pipeline is filled p. then subjects fully recognize the supply line and do not double order. While p=l prior experiment " showed P«l for many is 0, then P = 1, optimal, a subjects in a similar task (13). Expected demand EK represents each participant's forecast of incoming orders. Adaptive expectations are assumed: D^ = eDt.i + (i-e)D^t-i, o<e<i. (2) Consistent with behavioral decision theory (11, 12, 14), the rule utilizes information locally available to the decision maker and does not presume managers have global knowl- edge of the system structure or the cognitive capability The parameters (0, 44 the proposed rule subjects. A optimum performance. a, P, and S') were estimated for each participant by nonlinear least squares subject to the constraints power of to solve for is excellent: 0<6<1 and a, the mean R^ P, is S' >0 (table 2). 71%; R2 is The explanatory less than 50% for only 6 of large majority of the estimated parameters are significant, and the estimated parameters are systematically related to performance (10). Though the subjects' disequilibrium response is suboptimal, we expected that their decision rules would be stable and swiftly return the system to a low-cost equilibrium. To Sterman, Mosekilde, and Larsen test this hypothesis we 6 simulated the experimental system using each set of estimated Thirty parameter sets (68%) do indeed produce stable behavior. parameters (15). However, one periodic and three quasiperiodic solutions appear. Ten (23%) yield aperiodic (chaotic) behavior. Compare Figure 3 shows several characteristic simulations. = strange attractor produced by the parameters of subject 4 (0, a, P, S' S' = 0, 0.3, 0.15, 17). subject 36 (0, a, (3, produced by the (stable) To parameters for subject 27 values of .25 and 17, respectively, and varied The space chaotic solutions. The unstable (0, a, p, S' = 0.2, 0.3, 0.05, complex, including solutions arise when there each sector's immediate supplier, saler's inventory. Low is is and S' e.g., when frequencies arise when at representative stable, periodic, quasiperiodic, from the nonlinear coupling of several in the distribution chain. and oscil- These in the system. High sufficient inventory so that orders can be filled to wait for the factory to receive orders, Periodic solutions arise set 8). and P over the managerially meaningful coupled through the availability of inventory frequencies are produced frequencies a we by the multiple inventories and time delays latory feedbacks created oscillators are (figure 4) is 1.0, 0.65, 0.40, 15) to Note the long and apparently chaotic transient explore the structure of the parameter space, interval [0,1]. the shape of the by the retailer's orders aie filled out of the whole- when inventories are inadequate, forcing customers produce the beer, and finally ship it downstream. the parameters of the heuristic are such that the ratio of these rational; quasiperiodic solutions arise when this winding number is irrational, and chaotic solutions arise when the individual oscillators are sufficiently unstable (16). In general, larger values of a and smaller values of p are destabilizing (figure 4). To the extent a subject ignores the supply line of unfilled orders (P<1), a stock shortfall causes orders to be placed each period even after sufficient orders are in the supply line, leading to excess inventories and oscillation. stock adjustment (a>0.4) since However, the transitions Such overordering more is exacerbated by more aggressive will be ordered in response to a given stock shortfall. between modes are not monotonic. Note particularly the two narrow fjords of stable solutions which snake between the periodic and chaotic solutions. Sterman, Mosekilde, and Larsen 7 Figure 5 magnifies a region of parameter space unstable solutions. The at the transition oscillators staircase is complicated; however, by the fingers of stable solutions These regions of from for a=0.90. chaos. The structure of the which cut across the appear to have fractal boundaries as well. chaos also exists and also appears to be As P decreases from .444 to .434 the fractal. A transition Figure 6 varies P system moves irregularly from stability to Successive magnification reveals the self-similarity of the fractal boundary. It is the stability to stability from and form a devil's staircase (16). Further magnification reveals the characteristic fractal steps of the staircase. directly stable to alternating bands of periodic and quasiperiodic solutions arise mode-locking between the various staircase. from common in the social sciences to dynamics of human systems that formal rules are, if assume that not optimal, then decision-making behavior and thus at least stable. results show which characterize actual managerial decision making can produce an extraordinary range of disequilibrium dynamics, including chaos. number of These issues for social scientists. Such complexity raises a Policy interventions often imply changes in the parameters of a decision rule or model. How space" contains fractal basin boundaries? In such systems changes produce unpredictable qualitative changes circumstances which differ only slightly. can policy analysis be conducted in behavior. Do robust Experience if the "policy on the margin may may not transfer to principles of policy design exist in such systems? Even though deterministic cause-effect relations exist in chaotic systems, impossible to predict the effects of small changes in initial it is conditions or parameters. Does such 'randomness' slow the discovery of cause and effect by agents in the economy and thus hinder learning or evolution towards efficiency? Indeed, does learning alter the parameters of decision rules so that systems evolve towards or further away from development of theory and experiment are required assess the significance of chaos in human systems. the chaotic regime? to Much answer these questions and .. 8 Sterman, Mosekilde, and Larsen C. Grebogi, E. Ott, 1 J. A. Yorke, Science, 238, 632 (1987), and the references therein; R. May, Nature (London) 261, 459 (1976); Bristol, 1984); J. Thompson and H. P. Cvitanovic, M. Ed. Universality in Chaos (Hilger, Stewart, Eds. Nonlinear Dynamics and Chaos (Wiley, Chichester, 1986). R. Day, 2. Am. Econ. 3 (1986); J. M. Grandmont and J. M. Grandmont, Econometrica, 8, 397 (1987). E. Mosekilde M. J. Stutzer, /. Econ. 3. W. 4. A. Wolf, 5 O. Morgenstem, A. Brock, J. J. Malgrange, P. 53, 995 (1985); H. and Organization, 35, 80 (1988); . Rev., 72, 406; et al., W. J. Lorenz, Econ. Theory, 40, /. Econ. Behavior European J. of Operational Research, Dynamics and Control, 2, 353 (1980). Econ. Theory, 40, 168 (1986); P. Chen, System Dynamics Review, 4, 1988. Swift, H. Swinney, On the J. Vastano, Physica 16D, 285 (1985). Accuracy of Economic Observations, 2nd ed. (Princeton, Princeton, 1963). 6. J. D. Sterman, Management Science, 33, 1572 (1987). 7. J. W. Forrester, Industrial Dynamics (Cambridge, MIT, 1961); V. Zamowitz, Orders, Production, and Investment - Cyclical and Structural Analysis (New York, NBER, Dynamics, (Cambridge, MIT, 1963). 8 W. 9. C. Plott, Science, 232, 732 (1986); V. Smith, 10. J. E. Jarmain, Ed., Problems in Industrial D. Sterman, System Dynamics Review, 1973). American Economic Review, 72, 923 (1982). 148 (1988); 4, J. D. Sterman, Management Science, forthcoming. 1 1. A. Tversky, and D. Kahneman, Science, 185, 1124 (1974). 12. R. Hogarth, 13. J. Judgement and Choice, 2nd ed. 1987). D. Sterman, Organizational Behavior and Human Decision Processes, forthcoming 1988. 14. H. A. Simon, American Economic Review, 69, 493 (1979); R. Cyert and Behavioral Theory of the Firm (Englewood Management Science, 15. (New York, Wiley, 31, Cliffs, Prentice Hall, 1963); J. J. March, A Morecroft, 900 (1985). Analysis of variance showed no strong relations between a subject's position in the distribution chain and the estimated parameters. We therefore assume identical parameters for each sector. Sterman, Mosekilde, and Larsen reducing the dimension of the parameter space from 16 to 9 4. In the simulations orders are real- valued while in the experiment orders are integer-valued. The experiment is a difference equation system; a continuous-time version also produces chaos with plausible parameter values [E. 16. Mosekilde and E. Larsen, System Dynamics Review, M. H. Jensen, P. 4, 131, (1988)]. Bak, T. Bohr, Phys. Rev. A, 30, 1960 (1984). Sterman, Mosekilde, and Larsen Figure "Beer Distribution 1. Game" 10 board. Subjects carry out five steps each simulated week: 1. Receive inventory and advance shipping delays. 2. Fill orders. 3. Record inventory or backlog. 4. Advance 5. Place orders. make Subjects Figure the order slips. 2. their decisions in step 5; steps 1-4 handle bookkeeping and other mechanics. Typical experimental results. Top: customer orders increase from 4 to 8 cases/week week in 5. Middle: The resulting orders placed by subjects (from bottom to top, Retailer, Wholesaler, Distributor, Factory). inventory=Inventory-Backlog). Bottom: Effective inventory levels (Effective Tick marks on y-axes denote 10 amplification, and phase lag as the disturbance propagates Figure 3. showing Note the from customer oscillation, to factory. Simulation of the decision rule with estimated parameters. Top: phase portraits distributor inventory vs. factory inventory. transient; units. flow is Weeks 1000-1600 shown to remove clockwise. Bottom: The parameters for subject 27 produce a long chaotic transient (Retailer orders shown). Figure Parameter space of the simulated system. The stock and supply 4. parameters S'=17. a and P are varied over the interval [0,1] in line adjustment increments of .025, holding 0=.25 and Green=stable; red=periodic; blue=quasiperiodic; orange=chaotic. Note the two fjords of stable solutions separating regions of unstable behavior. Figure 5. Magnification of parameter space by a factor of 100 compared to figure 4. Green=stable; red=periodic; blue=quasiperiodic; yellow=period-doubling. The periodic H Sterman, Mosekilde, and Larsen solutions have rational winding number, indicating the high frequency oscillator locks into a rational multiple of the low frequency cycle. low frequency cycles completed before Figure 6. The numbers 3, 4, 5, etc. indicate the the system repeats. Transition from stability to chaos as P varies, holding a=.90. orange=chaotic. number of Green=stable; Successive magnification reveals the irregularity characteristic of a fractal boundary. Rightmost column of figure reflects magnification of 10^ compared to figure 4. 12 Sterman, Mosekilde, and Larsen Table 1. Summary of experimental results. Data are means of 11 trials. Actual costs exceed optimal levels by a factor of 10, a highly significant margin. Total COSTS $2028 Mean Benchmark* Ratio t-statistic: Hq: Mean cost = Benchmark Retailer Wholesaler Distributor Factory Sterman, Mosekilde, and Larsen Table 2. Summary of estimation results for 44 subjects. The ordering determines the speed of adjustment of the demand forecast; 1-2. response to inventory shortfalls; subjects; S' is the desired stock. Parameter: Mean 13 (std. dev.) Median: Minimum: Maximum: 6 (3 is heuristic is given a determines the the fraction of the supply line accounted for See (10) for by eqs. by the details. a P 0.36(0.35) 0.26(0.18) 0.34(0.31) 0.25 0.00 1.00 0.28 0.00 0.80 0.30 0.00 1.05 R^ S' 17 15 38 (9) 0.71 0.76 0.10 0.98 (0.22) Sterman, Mosekilde, and Larsen f 14 Sterman, Mosekilde, and Larsen 40 15 Customer Orders 10 30 Weeks Trial 5 -^ 10 UJ 20 Weeks 30 U Sterman, Mosekilde, and Larsen 30 '1 1 SutJiect 4 'O a -30 10 -10 -80 n Factory Inventory (cases -10 Factory Inventory (cases) -80 SuDiect 27 i 15 5: 10 o 5 6G0 Sterman, Mosekilde, and Larsen 17 a« Sterman, MDsekilde, and Larsen 8 0.085 !f —¥-*' 6 i * t 1 4 * ' -•------ P 7 * » 0.080 - ««<.4*«44««. 0.075 -fS 0.370 0.375 as 0.380 Sterman, Mosekilc3e, and Larsen Q.UL 0.U2U 0./.39 19 0.^387 0.43863 Sterman, Mosekilde, and Larsen 20 Appendix: Equations for Production-Distribution System The equations simulations (figure backlogs = are presented in the sequence of calculation used in the experiment and equilibrium conditions are: Initial 1). inventories = 12 units, Customer demand is initially 4 and The computer programs estimation are available from the Step The contents of 1: sector's inventory (I) are right (Dl) are same moved added into first for the simulations W, and parameter author. the shipping delay (Dl) immediately to the right of each The contents of to inventory. the shipping delay on the Dl. Factories advance the production delays Dip and D2p far in the fashion. Ij(t) = Ij(t-l) + Dlj(t-l) (Al) Dlj(t) = D2j(t-l) Step 2: Retailers examine Incoming Orders new week rises to 8 in Subscripts denote the sector of the distribution chain (Retailer R, Wholesaler Distributor D, and Factory F). all = 4 units (orders placed, incoming orders, production delays, 0, all other variables shipping delays, and expected orders). 5. all (A2) examine (10). All sectors B from orders plus any Backlog retailer's the top card on the Customer Order deck (CO); fill orders. The Shipment rate all S must equal the the prior period, to the extent inventory permits. shipments go directly to the customer and leave the system. others are placed in the shipping delay D2 Shipments of of the downstream sector (A4). others The all Shipments also reduce inventory (A5). SR(t) = Si(t) MIN(CO(t) -f = MIN(IOi.i(t-l) + B|(t-l),Ii(t)), D2k(t) = Sk^i(t), Ij(t) = Ij(t)-Sj(t) Step 3: (A3) BR(t-l), lR(t)) k= 1 = W,D,F (A3') R,W,D All sectors record inventory or backlog. (A4) (A5) • The net change ence between incoming orders and shipments. Effective Inventory EI is in backlog is the differ- inventory less backlog. Sterman, Mosekilde, and Larsen BrCD = BR(t- 1 ) + CO(t) 21 - SrCD (A6) Bi(t) = B,(t-l) + IO,.i(t-l)- S,(t), EIj(t) = The = W,D,F (A6') (A7) Ij(t)- Bj(t) Step 4: Each sector advances the orders. 1 week's Order slip containing last O to incoming factory puts last week's order in the top production delay D2p. I0,,(t) = 0j,(t-1), R,W,D k = (A8) D2F(t) = OF(t-l) Step supply line 5: is (A8') Each sector places the sum orders. First the Supply Lines are calculated. of units in the two shipping delays, the backlog of the supplier and orders placed the previous week. Since the factory line is SL is the primary producer, (if any) supply simply the contents of the production delays. SLk(t) = Dli.(t) + D2i,(t) + Bk^i(t) + I0j,(t), k= R,W,D (A9) SLF(t) = DlF{t) + D2F(t) Each its The sector's forecast of (A9') incoming orders (expected Demand) D^ is formed by adaptive expectations, with smoothing parameter 6. D-(t) = ejlOj(t- 1 ) + Finally, orders are the given between the desired Stock line. S' ( l-ej)D.(t- (AlO) 1 by demand forecast adjusted by a fraction a of the difference and the effective inventory, including a fraction p of the supply Orders must be nonnegative. Oj(t) = MAX(0,D-(t) + aj(S'j- Elj(t)- PjSL/t))) (All) Sterman, Mosekilde, and Larsen 22 Appendix. Estimated parameters, Beer Distribution Trial & Position R W D F R W D F R W D F R W D F R W D F R W D F R W D F R W D F R W D F 10 R W D F 11 R W D F e 0.90 a P S' Game r2 Mode u u u56 V^ ^b Date Due JUL JAN. 3 2 1 1 19a 199* 3 TQfiD DOS 37b ISM