«Pr HD28 .M414 ^<^%. »«s,, TsS;;-- NOV 12 1987 Ki'i^-B} ALFRED P. WORKING PAPER SLOAN SCHOOL OF MANAGEMENT DETERMINISTIC CHAOS IN MODELS OF HUMAN BEHAVIOR: METHODOLOGICAL ISSUES AND EXPERIMENTAL RESULTS JOHN D. STERMAN OCTOBER 1987 WP 1945-87 MASSACHUSETTS INSTITUTE OF TECHNOLOGY 50 MEMORIAL DRIVE CAMBRIDGE, MASSACHUSETTS 02139 r '^ DETERMINISTIC CHAOS IN MODELS OF HUMAN BEHAVIOR: METHODOLOGICAL ISSUES AND EXPERIMENTAL RESULTS JOHN D. STERMAN OCTOBER 1987 WP 1945-87 D-3921 Deterministic Chaos in Models of Human Behavior: Methodological Issues and Experimental Results John D. Sterman Associate Professor Sloan School of Management Massachusetts Institute of Technology Cambridge, 02139 MA Forthcoming in the Special Issue on Instabilities and Chaos System Dynamics Review October 1987 D-3921 ABSTRACT Recent work has shown literature contain previously that several well-known models model of the Two wave and of the most the production- Beer Distribution Game. The significance of these theoretical developments hinges on whether the chaotic regimes space. system dynamics unsuspected regimes of deterministic chaos. extensively analyzed are Sterman's model of the economic long distribution in the lie in the realistic region of parameter Further there are major questions regarding the descriptive accuracy of models of human systems which exhibit chaos. experiments mean empirical studies questions. An Data limitations and the at the inability to conduct controlled aggregate level are not likely to resolve these alternative approach is based on laboratory experiments provide a simulated environment for the study of human decisionmaking in which models behavior. Recently laboratory experiments have been conducted to analyze decisionmaking behavior in the long wave model and shows the Beer Distribution Game. This paper describes the experiments and that the behavior of the subjects is explained well with a simple heuristic long used in system dynamics modeling and well grounded in behavioral decision theory. The parameters of the proposed decision rule are estimated econometrically for each subject. The parameters which characterize a significant minority of the subjects are direct experimental evidence that chaos can be real shown to produce chaos. This produced by the decisionmaking behavior of people has important implications for the formulation, analysis, and testing of models of human behavior. D-3921 Recent work has shown literature contain previously well-known models that several in the system dynamics unsuspected regimes of deterministic chaos. The work of Day (1982a, 1982b) provides an early example of chaotic behavior in economic models, while Mosekilde and others (1987) have developed corporate models which exhibit chaos. the most extensively analyzed such models are Sterman's Two of model of the economic long wave (Sterman 1985, 1986, Rasmussen, Mosekilde, and Sterman 1985) and the productiondistribution system or Beer Distribution Mosekilde and Larsen, produced in these this issue). systems results hinges in large is Game While (Forrester 1961, Jarmain 1963, Sterman 1984, the demonstration that chaos can be endogenously an important theoretical development, the significance of the measure on whether the chaotic regimes lie in the realistic parameter space or whether they are mathematical curiosities never observed region of in the real Further there are major questions regarding the descriptive accuracy of the decision system. rules postulated in the attractors. The models of human systems developed practical significance of chaos in policy-oriented modeling remains unclear to date which contain strange and other phenomena such until it as self-organization can be determined that these phenomena can occur in models whose decision rules are grounded in empirical study of the actual decision processes of the agents. It is difficult if not appeal to the aggregate empirical data. been at most five reliable results. wave, for example, there have industrial revolution, too few for statistically Worse, the data required are simply unavailable, and much of it is corrupted this issue). alternative approach is based on laboratory experiments simulated environment for the study of the long In the case of the long long-wave cycles since the by measurement error (Chen, An impossible to resolve such issues by wave model and Beer in which models provide human decisionmaking (Sterman Distribution Game 1987a). Recently, (traditionally used as a teaching rather than as a research tool) have been analyzed as experiments in dynamic decisionmaking a D-3921 These experiments were conducted primarily (Sterman 1987b, 1987c). to study the heuristics with which people manage a complex dynamic environment. In both experiments the decision task of the subject was manage to a stock in the face of losses, delays in acquiring new units, multiple feedbacks and other environmental disturbances. This paper describes the experiments and shows that decisionmaking behavior two experiments is The departures from optimal behavior significantly suboptimal. common systematic, suggesting subjects use heuristics for stock in the are managment. The experiments show that the behavior of the subjects can be modeled well with a simple stock- management behavioral decision theory. The parameters of econometrically for each subject. parameters. shown to system dynamics modeling and well grounded heuristic long used in the The systems The parameters which in proposed heuristic are estimated are then simulated with the estimated characterize a significant minority of the subjects are produce chaos. This direct experimental evidence that chaos can be produced by the decisionmaking behavior of real people has important implications for the formulation, analysis, and testing of models of human behavior. The Stock Management Problem The regulation of decisionmaking tasks. a stock or The manager seeks within an acceptable range. altering the stock's inflow as to system and outflow the initiation of a control action and the stock and the perception of that may common dynamic to maintain a quantity at a target level, or at least rates. Typically a manager must set the inflow rate so compensate for losses from the stock and vary and one of the most Stocks cannot be controlled directly but rather are influenced by which may push the stock away from may state is its its to counteract desired value. effect environmental disturbances Frequently there are lags between on the stock, and/or lags between a change change by the decisionmaker. be influenced by the manager's own actions. The duration of in these lags D-3921 Stock management problems occur macro. at many levels of aggregation from the At the level of a firm, managers must order parts and raw materials so micro to the as to maintain inventories sufficient for production to proceed at the desired rate, yet prevent costly inventories from piling up. They must adjust for variations in the usage and wastage of these materials and for changes in their delivery delays. At the level of the individual, people down regulate the temperature of the water in their morning shower, guide their cars highway, and manage their checking account balances. At the macroeconomic Federal Reserve seeks to manage the stock of economic growth while avoiding inflation The generic stock management control problem Considering first Si= ' I may A in credit demand. be divided into two parts: the decision rule used by the manager the stock and flow structure, the stock of interest S of the acquisition rate ' (ii) level, the so as to provide sufficient credit for and compensating for variations stock and flow structure of the system; and 1). money is ^ X acquisition rate will (1) itself, and may also depend on other and exogenous variables U: (2) depend on the supply line SL of units which have been not yet received, and the average acquisition lag X. In general, = The supply (figure the accumulation Lt = fL(St,Xt,Ut). At itself X may be ordered but a function of the and the other endogenous and exogenous variables: fA(SLat). line is (3) simply the accumulation of the orders which have been placed O less those which have been delivered: SL,= the (A,-L^)dx+S, K endogenous variables supply line (i) less the loss rate L: Losses from the stock must depend on the stock The the I •^0 (O,- A,)dx+SL, ° (4) D-3921 The structure represented by figure that the functions arbitrarily and 1 eq. (1-4) is There quite general. governing losses and the acquisition lag are complex feedbacks among the endogenous no presumption There may be linear. variables, is and the system may be influenced by numerous exogenous forces including noise and nonstationarity of the underlying equilibrium. Consistent with the behavioral foundations of system dynamics modeling (Morecroft 1983, 1985, Sterman 1987a), behavioral decision theory (Hogarth 1987, Tversky and Kahneman 1974), and the theory of bounded rationality (Simon 1979, Cyert and March 1963), the proposed decision rule utilizes information locally available to the decisionmaker not presume the manager has global knowledge of the structure of the system. decision rule recognizes three motives for ordering which any stock and does The generic management heuristic must include: Order enough to (1) replace expected losses from the stock, (2) reduce the discrepancy between the desired and actual stock, and (3) maintain an adequate .supply line of unfilled orders. 1. when Replacement of losses. The replacement motive the desired is straightforward. and actual stock are equal, the manager must continue replace ongoing losses. Losses may arise from usage (e.g. would cause 2. demand a the stock to fall Stock adjustment. mechanism below the desired to order enough to shipments from an inventory of finished goods) or decay (e.g. the depreciation of plant and equipment). losses In equilibrium, Failure to replace level, creating steady-state error. Errors in forecasting losses or changes in the desired stock to adjust orders above or below replacement. Orders to reduce the discrepancy between the desired and actual stock form a negative feedback loop which regulates the stock (shown in the bottom part of figure 1). Any rule compensate for discrepancies between the desired and actual stock at all. Such rules could not respond to a change which fails to fails to control the in the desired stock, stock nor restore the stock to ^ D-3921 The stock would follow the desired value if displaced. a random walk as the system is bombarded by shocks. Supply 3. give stock Delays between the line adjustment. management systems significant inertia to ensure a stable response to shocks. initiation and impact of control actions and should be accounted for by managers Failure to account for the supply line results in overcorrection and instability. Consider cooking dinner on an electric range. If one turns the heat down comes just as the pot continue to heat the pot, boiling to a boil the supply line of heat in the coils of the range will it over and ruining dinner. The following equations formalize in most real life situations the ordering heuristic proposed above. First, orders must be nonnegative, Ot = MAX(0,IOt) where 10 is The (5) the indicated order rate, the rate indicated by other pressures. indicated order rate is based on the anchoring and adjustment heuristic (Tversky and Kahneman 1974). Anchoring and adjustment an unknown quantity estimated by is first recalling a then adjusting for the effects of other factors which obscure. a is common known may Anchoring and adjustment has been shown judgmental strategy in which reference point (the anchor) and be less salient or whose effects are to apply to a wide variety of decisionmaking tasks (Einhom and Hogarth 1985, Davis, Hoch, and Ragsdale 1986, Johnson and Schkade 1987, Lopes 1981, Hines 1987, Sterman 1987dX Here the anchor expected loss rate L^. Adjustments are then made desired and actual stock AS in is the response to discrepancies between the and between the desired and actual supply line ASL: lOt = L^t + ASt + ASLt. The expected loss rate may (6) be formed in various ways. economics and management science include static Common assumptions expectations L^t = L* (a constant or in D-3921 equilibrium value), regressive expectations L^t= yLt-l + (l-Y)L*, 0<y<l, adaptive expectations L«t= 6Lt-i + (l-6)L^t-l, 0<9<1, and extrapolative expectations, AL^t = lcoi*ALt-i, where A is The adjustment regulates the stock. the first-difference operator and coi>0. for the stock The proposed linear in the discrepancy AS creates the chief negative feedback loop heuristic assumes for simplicity which adjustment that the is between the desired stock S* and the actual stock: ASt = as(S*t-St) (7) where the stock adjustment parameter as is the fraction of the discrepancy ordered each is formulated analogously as period. The adjustment for the supply line ASLt = asL(SL*twhere SL* line. O* is SLt) (8) the desired supply line and The desired supply asL is the fractional adjustment rate for the supply line in general is not constant but depends on the desired throughput and the expected lag between ordering and acquisition of goods: SL*t = Xh*0*t. The adjustment (9) overordering and also compensates for changes in the acquisition lag X. example, ASL which avoids for the supply line creates a second negative feedback loop induces sufficient additional orders to restore possible representations for X^ and If A=0*. There X rises, for are a varier>' of O*, ranging from constants through sophisticated forecasts. Experiment The experimental The Economic Long Wave protocol for the long Meadows 1985 and Sterman economics are discussed I: wave model is described in Sterman and 1987a, 1987b). The methodological foundations of experimental in the seminal work of Smith (1982) and Plott (1986). The model represents the aggregate capital-producing sector of the economy (figure 2). The dynamic D-3921 hypothesis behind the long the process by wave can be stated in about 15-25 years in First, Second, in the orders and acquires capital from damped due to construction lags, demand is inherently fluctuations with a period of aggregate the capital-producing itself. This multiplier effect or which destabilizes the ordering' creates a positive feedback loop individual firms, changing the to would produce damped response to a shock. economy sector of the parts. which individual firms adjust production capacity In isolation an individual firm oscillatory. two 'self- oscillatory tendencies of oscillation to a limit cycle with a 40-60 year period. Simulation and formal analysis confirm the dynamic hypothesis (Sterman 1985; Rasmussen, Mosekilde and Sterman 1985). As the self-ordering loop becomes stronger, damping drops The system goes through rapidly. a Hopf bifurcation and produces a limit cycle. Further increases in the strength of self-ordering proceed through period doublings and ultimately to chaos. model In the experiment the is transformed into a capital-producing sector of a simple economy. demand for capital for capital arise orders. by adjusting its in which subjects manage meet desired production. Orders from the exogenous consumer goods sector and from essential structure the game reflect a simplified and dynamics. the Subjects must balance the supply of and their production capacity to The equations underlying preserve game In the game the subject's own form of the original model but decisions are made in discrete time intervals representing two years, and the model becomes a third-order difference equation system. The equations of production DP the game are given below. or Production Capacity PC. Capacity capital/output ratio k is Production is PR is the lesser of desired proportional to the capital stock. The one period (two years): PRt = MIN(DPt,PCt) (10) PCt = (11) Kt/K. D-3921 The capital stock of the capital sector depreciation CD. Depreciation capital stock x The is is is augmented by acquisitions proportional to the stock. AK and diminished by The average lifetime of the 10 periods (20 years): Kt+i = Kt + (AKt-CDt) (12) CDt = (13) Kt/t. acquisition of capital by both the capital and goods sectors (AK and AG) depends on supply line of unfilled orders each has accumulated (the backlogs BK and BG) and the fraction of demand satisfied FDS. Each period both the capital and goods sectors acquire the supply line of unfilled orders unless the capital sector amount. In this acquisition lag case acquisition by each sector X is thus 1/FDS, and is AK+AG=PR is the full unable to produce the required reduced in proportion to the shortfall. at all times, ensuring that output The is conserved: AKt = BKt*FDSt (14) AGt = BGt*FDSt (15) FDSt = PRt/DPt (16) DPt = BGt + BKt. (17) The supply lines of unfilled orders for BG and BK are augmented by orders for by each sector and emptied when those orders are delivered: capital placed BKt+i = BKt + (OKt New each sector AKt) (18) BGt+i = BGt + (OGt-AGt). (19) - orders placed by the goods sector are the exogenous input to the system to which the subject of the experiment for their own must respond by choosing an appropriate amount of capital to order use: OGi = exogenous (20) OKt = (21) determined by subject. Subjects are responsible for only one decision - how much capital to order. The goal D-3921 9 of the subjects in making these decisions score minimize to is their total score for the trial. defined as the average absolute deviation between desired production is production capacity PC over the T DP The and periods of the experiment: T S = (1/T) S I DPt - PCt (22) |. t=0 The score indicates how equally for both excess The experiment well subjects balance demand and excess is implemented on demand and supply. Subjects are penalized supply. IBM PC-type microcomputers. A 'game board' is displayed on the screen providing subjects with perfect information. Color graphics and animation highlight the flows of orders, production, and shipments to increase the transparency of the structure. The conversion of the original model to a form suitable for experimental testing is described in Sterman 1987a. The subject population (N=49) consisted of students in MIT undergraduate, master's and doctoral management and engineering, many with extensive exposure contro^ theory; scientists and economists from various institutions in the to economics and US, Europe, and the Soviet Union; and business executives experienced in capital investment decisions including several company presidents and CEOs. Typical experimental results are shown in figure period 3 there is a one-time step input of 10% 3. All trials begin in equilibrium. in the orders of the goods sector. In The optimal response returns the system to equilibrium within 6 periods, producing a score of 19. contrast the vast majority of subjects produced significant oscillations (table The average 1). In score for the sample was 591; the lowest was 77. Next the proposed stock management heuristic (equations 5-9) was tested against the ordering behavior of the subjects. straightforward. The stock to be Adapting the heuristic to the experimental context managed is the capital stock K. The desired stock is is 10 D-3921 The proportional to desired production. stock CD. The DSL desired supply line loss rate L is simply the depreciation of the capital was specified according to equation 9, with the 0*= CD. expected acquisition lag X^ =X= additive disturbance term the proposed ordering rule for capital investment OKt = E, 1/FDS and desired throughput MAX(0,CDt + ASt + ASLt + Allowing an becomes: (23) Et) ASt = as(DKt-Kt) ASLt = asUDSLt - (24) BGt) (25) ' DSLt = ?.e*CDt= (l/FDSt)*CDt. To test the rule (26) only the two adjustment parameters as and asL need be estimated. All other Maximum likelihood data required to determine orders are presented directly to the subjects. estimates of the parameters for each trial Assuming subject to the constraints as, asL normally distributed then the maximum ^0. were found by grid search of the parameter space, the errors £ are independent, identical, and likelihood estimates of such nonlinear functions are given by the parameters which minimize the sum of squared Such estimates errors. are consistent and asymptotically efficient, and the usual measures of significance such as the test are t- asymptotically valid. Estimates for 49 trials together with t-statistics are given in table ability to explain the ordering decisions of the subjects is excellent. R2 2. The model's varies between 33% and 99+%, with an overall R^ for the pooled sample of 85%.2 All but two of the estimated capital stock adjustment parameters are highly significant. parameter is The supply line adjustment significant in 22 trials. Sterman 1987b analyzes the estimation results and identifies several 'misperceptions of feedback' which are responsible for the subjects' poor performance. tendency for subjects to give insufficient attention to the supply line, One of these is the causing subjects to D-3921 11 continue ordering even after the construction pipeline contains sufficient units to correct any stock discrepancy. The present concern, however, is the relationship The estimated parameters parameters and the regimes of behavior in the model. making behavior of the decision When actual people. between the estimated characterize simulated in the model are the estimated decisoin rules of the subjects inherently stable, or do they produce limit cycles, period multiples, or chaos? Table 2 also indicates the mode of behavior produced by simulation of the decision rule The parameters estimated with the estimated parameters. stable. Most of these produce overdamped behavior of for thirty subjects Seven parameter multiples. The parameters which sets are the capital stock in response to the produce limit cycles of period step input. (61%) characterize ten subjects 1, and 2 produce period (20%) produce chaos. Inspection of table 2 shows that the subjects whose parameters are stable performed best in the task while those whose parameters produce limit cycles and chaos generally had the highest scores. Figure 4 shows the phase portrait for several of the more interesting parameter Note the period 5 cycle produced by (as, asL) = produced by (.42, 0) is, upon closer inspection, seen Figure 5 locates the modes of behavior intuition and formal values of asL. (.46, .33). in stability analysis, stability is More The apparent period sets. 2 cycle to be chaotic. parameter space. Consistent with both enhanced by small values of as and large aggressive attempts to correct the discrepancy between the desired and actual capital stock are destabilizing: by ordering more aggressively the self-ordering loop causes a larger increase in total demand, thus exacerbating disequilibrium and encouraging still is larger orders in future periods. stabilizing Conversely, more aggressive response to the supply line by constraining orders as the supply line fills. 1^ D-3921 the periodic and chaotic attractors tend to occur for large values of Thus values of asL- Yet the route to chaos is apparently way Stability gives chaos as as increases for asL=0. 2 at (.42, 0), and ultimately to chaos for as ^.55. slightly chaotic period 2 are period lies 1 limit cycles. results performance in a to period one the route to limit cycles, then to period Yet the parameters which produce the by parameters which produce surrounded on the ray asL = Similarly, the period 5 attractor produced by the parameters (.46, .33) between the region of The more complex. Consider as and small show stability and the region of period that the parameters common and 1 behavior. which characterize actual decisionmaking important macroeconomic situation can produce a wide range of behavior, including chaos. Experiment The "Beer Distribution II: The Production-Distribution System Game" and distribution system developed at is a role-playing simulation of an industrial production MIT to introduce students of management to the concepts of economic dynamics and computer simulation. In use for neariy three decades, the game has been played all over the worid by thousands of people ranging from high school students to chief executive officers and government officials. The game is played on a board which portrays in a simplified fashion the production and distribution of beer (figure 6). Orders for and cases of beer are represented by markers and pennies which are physically manipulated by the players as the game proceeds. Each brewery consists of four sectors: One person manages each sector. Customers demand beer from the retailer in turn orders beer wholesaler's inventory. retailer, wholesaler, distributor, and factory (R, Customer demand retailer, who is W, D, F). represented on a deck of cards. ships the beer requested out of inventory. from the wholesaler, who ships the beer requested out of the Likewise the wholesaler orders and receives beer from the The D-3921 13 who distributor, and receives beer from the in turn orders factor^'. The factory produces the At each stage there are shipping delays and order receiving delays. These represent beer. the time required to receive, process, ship, and deliver orders, and as will be seen play a crucial role in the dynamics. The subjects' objective is to minimize cumulative team costs over the length of the game. Inventory holding costs for each sector are $.50 per case per week, and stockout costs (costs for having a backlog of unfilled orders) are $1.00 per case per The problem. decision task of each subject a clear example of the stock management Subjects must keep their inventory as low as possible while avoiding backlogs and satisfying customer The is week. demand. Inventory cannot be controlled lag in receiving beer directly but must be ordered. potentially variable: if the wholesaler has beer sufficient to cover is the retailer's orders the retailer will receive the beer desired after three weeks. But if the wholesaler has run out, the retailer must wait until the wholesaler can receive additional beer from the distributor. Only the acquiring inventory (there The protocol in equilibrium. week. To is no limit to the production capacity of the factory). for the experiment The is described in Sterman 1987c. Each inventory contains 12 disturb the system customer and remains cases. demand Customer demand The game is initially increases to eight cases per is initialized four cases per week in week 5 at that level thereafter. results reported here A over a period of four years. consistency. for factory, the primary producer, faces a constant delay in more than Trials in a were drawn from four dozen games (192 subjects) collected computer model of the game was used to test the records for which there were accounting errors of more than a few cases per week few weeks in any of the four sectors were discarded from further analysis. Eleven games were retained, providing 44 subjects.^ That sample consists of undergraduate, MBA, and Ph.D. students at MIT's Sloan School of Management, executives from a variety of D-3921 14 firms participating in short courses on computer simulation, and senior executives of a major computer firm. Typical experimental results are shown in figure 7 and summarized in table results show 3. The from optimal. The average team cost the behavior of the subjects to be far is Like the macroeconomic experiment, the results exhibit ten times greater than optimal. several strong regularities. 1. Oscillation: The trials are all pattern of orders and of inventory is characterized by instability and oscillation. The dominated by a large amplitude fluctuation with an average period of 21 weeks. 2. Amplification: The amplitude and variance of orders increases steadily as one moves from customer to retailer to factory. The peak order average more than double the peak order rate generated increase from 4 to 8 cases per week. factor)', the 3. By rate at the factory level is at the retail level. on Customer orders the time the disturbance has propagated to the order rate averages a peak of 32 cases. Phase to the factory. lag: The peak order The phase lag is rate tends to occur later as one moves from the retailer not surprising since the disturbance in customer orders must propagate through decisionmaking and order delays from retailer to wholesaler and so on. Next the proposed decision game and cast in a form rule must be adapted to the particular situation in the beer suitable for estimation of the parameters. to the inventory of the subject and the supply line SL to the sum The stock S corresponds of orders in the mail delays, the backlog of the subject's supplier (if any), and the beer in the shipping delays. rate is the rate at which each subject receives orders. To test the rule it is The loss necessary to specify expected losses L^, the desired stock S*, and the desired supply line SL*. Expected losses from the stock are the immediate customer to place orders, that is, rate at which each subject expects the retailer's forecast of the their customer order rate, D-3921 15 the factory's forecast of the distributor's order rate, etc. postulated. Adaptive expectations are widely used economic systems, they are often a in Adaptive expectations are simulation modeling of corporate and good model of the evolution of expectations in the aggregate (Sterman 1987d, Frankel and Froot 1987), and they are one of the simplest formulations for expectations flexible enough to adapt to a nonstationary process. the subjects lack the information the desired stock and time S* and desired supply Because to determine optimal inventory or supply line levels line SL* are both assumed to be constants which must be estimated. The generic decision rule of eq. (5-9) thus becomes: Ot = MAX(0,IOt) (27) IOt=Let + ASt + ASLt (28) Let = eLt-l + (l-e)Lei.i. ASt = as(S* - 0<e<l (29) St) ASLt = asL(SL* - (30) SLt). (31) Defining P = asUct-s and S' = S* + PSL* yields lOt =Let + as(S'-St-pSLt). Note that since S*, place more emphasis on SL*, asL and as are (32) all >0, S'>0. Further, the supply line than on the inventor)' it is unlikely that subjects will itself: the supply line does not directly enter the cost function nor is it asLS can be interpreted as the fraction of the supply line taken as, meaning 0<(3<1. Thus into account by the subjects. If (3 P= 1, as salient as inventory. Therefore it is probable that then the subjects fully recognize the supply line and do not double order. If P = 0, then orders placed are forgotten until they arrive, encouraging overordering and instabilitj.'^ In contrast to the macroeconomic experiment, the stock management decision rule as D-3921 16 implemented in the beer game contains four parameters to be estimated (6, as, S', and p). The parameters were estimated by An additive disturbance term is the same procedure used in the long wave experiment. assumed, and the disturbances are assumed to be normal. Estimates for each sector of each trial i.i.d. were found by grid search of the parameter space subject to the constraints 0<6<1 and as, S', p >0.5 The estimated parameters than 50% for only 6 of significant. As 44 subjects.^ in the long A large wave experiment decisionmaking behavior of the subjects. the decision rule Sterman 1987c It is feedback structure which cause oscillation in the Next given far too the beer trial is less is an excellent model of the relates the estimated parameters to macroeconomic experiment are responsible game. Specifically, the supply line of weight by most subjects. game was simulated using estimated parameters. Each parameters. little 71%; R^ noteworthy that the same misperceptions of for the poor performance of the subjects in the beer is is majority of the estimated parameters are the dysfunctional performance in the task. unfilled orders The mean R2 are displayed in table 4. the stock adjustment rule with the involves four subjects, each of In the simulations, however, the whom could have different same parameters were assumed for each sector, vastly simplifying the analysis.*^ Table 4 indicates the one parameter 1, sets mode (70%) produce of behavior produced by each set of parameters. Thirtystable behavior. period 5, and period 12. Ten yield chaos. behavior produced by the system projections of phase space. is One each produces Since the beer game is limit cycles of period a 19th order exceedingly complex and there are many system the possible Figure 8 shows several of the more interesting phase portraits. Note, for example, the different shape of the attractor produced by simulation with the 17 D-3921 parameters of the Bassbeer Factory HeinekenS Factory Factory (9, as, p, as, P, S' = 0, (9, S' = .25, 1 The . as, P, S' = 1, .65, .40, 15) The retail sector difference in mode is compared The parameters estimated .3, .15, 17). 18) reveal a cycle of period 12 .3, 0, plotted against factory inventory. limit cycle of period (9, when to those of the for the Coors distributor inventory is of the same team, however, produces a probably due to the integer constraint on orders. The Figure 9 shows the distribution of the modes in parameter space. between the adjustment parameters and macroeconomic experiment, and for similar reasons. smaller values of P are destabilizing. line of unfilled orders even stability are similar to those That is, apparent in the In general, larger values of since more stock. (P=0) a stock discrepancy will cause orders to be placed each period Such overordering is and exacerbated by more aggressive stock adjustment (as>>0) will be ordered in response to a given discrepancy between the desired and actual Figure 9 also indicates that in general the smaller the desired inventory S' the less stable the system. Again, intuition readily explains the retailer to an unanticipated increase in customer lags in receiving beer. The the stock adjustment term. (S'»0) incoming The delay why this is so. If the orders can demand. Retail inventory all wholesaler has adequate inventory to be filled and the still orders swell the supply line in a vain attempt to replenish inventory. retailer finds inventory fill retailer finds orders drops falls due to the wholesaler through retailer's acquisition lag in receiving beer lengthens, the retailer's stock does arrive, the Consider the response of retailer will place additional orders with the however, the wholesaler runs out of inventory the finally as and to the extent the subjects ignore the supply after sufficient orders are in the supply line, leading to excess inventories oscillation. If, relationships incoming orders remains constant. placed further, When and fail to arrive. still more sufficient beer overshooting the desired level by huge margins 18 D-3921 (figure 7). S' plays such a crucial role in determining the stability of the system because determines how it close the system operates to the fundamental nonlinear contraint that shipments can only be made if inventor)' exists. Small values of S' make it more likely the system will run out of inventory thus entering the unstable region in which the acquisition lag X is thus long and variable. more stable Those subjects who maintained on average than those who attempted larger target inventory levels to cut costs by reducing were their buffer stock. There appears the system. to be no simple relationship between the parameter 6 and the modes of 6 controls the speed with which the adaptive forecast of incoming orders Faster adjustment of the adjusted and has two opposing effects on stability. (9=1) increases the gain between orders placed and incoming orders. demand is forecast Shocks are transmitted up the distribution chain more readily, destabilizing the system by causing larger inventory excursions upstream in response to a given change in orders downstream. On the other hand, sluggish adjustment (0=0) means variations in incoming orders are not quickly countered by the replacement term of the ordering rule, causing inventory to decline further than it would if replacement of incoming orders was swift. As a result, inventories may be exhausted, moving the system into the less stable zone in which backlogs build up and the acquisition lag lengthens. As behavior is in the macroeconomic experiment the results show an unexpectedly rich array of generated by the interaction of the decision processes of real people with the feedback structure of the environment. Discussion The discovery of nonlinear phenomena such as deterministic chaos in the physical world naturally motivates the search for similar behavior the social scientist faces difficulties in that search in the world of human behavior. Yet which do not plague the physicist, at least D-3921 19 not to the exist for same degree. Aggregate data many of most important the sufficient for strong empirical tests simply social systems. from the environment. The huge temporal and Social systems are not easily isolated spatial scales of these systems, vast number of individual actors, considerations of cost and ethical concerns on the systems themselves difficult at best. do not make Finally, the laws of controlled experiments human behavior are not as stable as the laws of physics. The social scientist is thus left with relevance of nonlinear systems may be phenomena in the modes in a in the quest to understand the domain of human behavior. Formal models of constructed and analyzed. existence of chaotic two main options social Previous work of this type has demonstrated the wide array of social and economic models. However, analysis of these models frequently showed that the chaotic regime lay far outside the plausible region of parameter space. the Worse, without empirical models was open phenomena remained to doubt. In the absence of empirical data the relevance of nonlinear is to develop laboratory experiments with simulated social Since experiments on actual firms and national economies are infeasible, simulation models of these systems must be used people. of the decision rules the verisimilitude of questionable. The second approach systems. tests to explore the Such experiments create 'microworlds' institutional structures, information, decisionmaking heuristics of real which the subjects face physical and in and incentives which mimic (albeit in a simplified fashion) those of the real world. The experiments described here presented subjects with a straightforward stock- adjustment task. Results show that the subjects' behavior can be modeled with a high degree of accuracy by a decision rule long used is in system dynamics models. The decision rule consistent with a vast body of empirical knowledge developed in behavioral decision theory, organizational studies, and psychology. The stock-adjustment heuristic appears to 20 D-3921 be robust: the feedback structure, information available, loss function, time available, "cover story" and many other dimensions of the experimental environment differed markedly. Performance in both settings was decidedly suboptimal. feedback structure of the system are apparent in The same misperceptions of the both cases. The experiments reported here demonstrate that chaos can in fact be produced by the decisionmaking processes of real people. The results are significant in several respects. The demonstration that chaos can be produced by the decisionmaking heuristics people actually Chaos thus appears use strengthens the argument for the universality of these phenomena. to be a common mode of behavior not only in physical systems but in social and economic systems as well. These results do not show States has in fact been chaotic. be settled. peril. But the Further empirical work results suggest Models of economic and evolution of that, say, the GNP required before such questions can is modelers can ignore nonlinear dynamics only social all times or that adjustment processes are stable. formulated so that they are robust in extreme conditions, since it is These principles have long been central Forrester and Senge 1980, Richardson and to economy is in or Models should be the nonlinearities necessarily introduced by robust formulations that crucially determine the system. at their dynamics should portray the processes by which disequilibrium conditions are created and dissipated, not assume that the near equilibrium at United in the modes of the system dynamics (Forrester 1961, Pugh 1981) and behavioral economics (Da)' 1984) and are important even for systems which do not contain limit cycles or strange attractors (Sterman 1985b). At the same time a number of questions regarding and other nonlinear phenomena in social the practical significance of chaos system modeling remain unresolved. Real social systems are bombarded by broadband noise, and it is well known that such random shocks severely degrade the point-predictability of most systems (Forrester 1961, App. K). Does 21 D-3921 the existence of stochastic shocks How swamp does the existence of chaotic regimes the predictability of that response? the uncertainty in trajectories caused by chaos? model influence in a Particularly troubling here the simulated beer response to policies, and the apparent nonuniformity Mosekilde and Larsen's of the distribution of modes in parameter space. bombing experiments with is its game show (this issue) carpet that there are islands of chaos in parameter space, and that the boundary between the periodic and chaotic solutions a finding consistent with analysis of such classic systems as Buffing's equation and Stewart 1986). The response of stability to may parameter variations Similarly, chaos frames, but all many et al. a steady-state such systems is a major area for future research. phenomenon which manifests over very long time much 1986, Weidlich et al. learning and evolutionary pressures. this issue). How Over such extended time horizons static decision functions remain constant throughout. here, for example, Yet there that subjects begin to learn within just a is the behavior of The parameters of the few cycles, modifying the parameters at least, the chaotic region into the region of stability. in policy-oriented show evidence (Sterman 1987b, of their ordering function. In the macroeconomic experiment, chaos and other nonlinear phenomena the but will themselves evolve with systems simulated for thousands of years and weeks, respectively. away from (e.g. does learning influence chaotic systems? The macroeconomic and corporate experiments described subjects dynamics shorter than those used in the analysis of chaotic parameters of the system cannot be considered Bakken 1987) The dev- policy-oriented models are concerned with transient dynamics and nearly with time horizons Mosekilde is in (Thompson not be monotonic or even predictable in such systems, raising major questions for policy analysis. elopment of principles for policy design is fractal, The learning moves the practical significance of models of social and economic behavior remains clouded while these questions are unanswered. More important, the results demonstrate the feasibility of subjecting theories of 22 D-3921 human behavior to experimental Many models test. of human systems may be built which produce an astounding variety of behaviors. Unfortunately many of these models are formulated without regard to the vast body of empirical knowledge documenting the decision processes of individuals. For example, rational expectations models economics presume in agents have perfect knowledge of and the cognitive capability to solve the system of equations which characterize the economy (Lucas 1976, Shaw strongly contradicted by evidence at all levels of aggregation 1984). These assumptions are (Simon 1984, Hogarth and Reder 1986) and form a poor foundation for descriptive models of economic dynamics. Similarly, physical scientists frequentiy specify decision processes in behavior by analogy with natural systems (Beylich 1986, Zeeman 1977, models of human Zimmerman 1986), or simply assert that the processes studied in natural systems extend by analogy to social dynamics (Ruelle 1984, May 1976). The natural sciences can indeed be a fruitful source of metaphor for theories of human behavior, and there of nonlinear science to social systems empirical grounding in tiie (e.g. is encouraging progress Prigogine and Sanglier, 1987). psychology of choice behavioral empirical test. as accepted But without scientists will continue to such models as ad hoc and continue to question their relevance. assess the validity of such in the application It a firm view has been difficult to models because the assumed decision rules were not subject In the future experimental tests of models of and commonplace as they are to human behavior should become in the natural sciences. Over time the application of these methods should identify the processes of judgment people actually use in dynamic decisionmaking tasks and which are acceptable as the basis for accurate modeling of human affairs. Models formulated on the basis of with the tools of simulation and offer the best systems. hope to modem knowledge of individual decisionmaking, analyzed nonlinear science, and subjected to experimental improve our understanding of the dynamics and evolution of social test D-3921 23 Notes 1. Order cancellations are sometimes possible and may exceed new orders conditions (e.g. the U.S. nuclear power likely to be subject to different costs industry in the 1970s). in extreme Since cancellations are and administrative procedures than new orders they should be represented separately as a distinct outflow from the supply line of unfilled orders rather than as negative orders. 2. Note that the function OK=f(-) does not contain an estimated regression constant. Thus the correspondence of the estimated and actual capital orders, not just their variation around mean values, provides an important measure of the model's explanatory power. Since the residuals e need not satisfy Zet = The alternative R2 = 1 - Set^ / lOKt^ is the conventional used (Judge et al. R^ 1980). is not appropriate. This R2 can be interpreted as the fraction of the variation in capital orders around zero explained by the model. 3. Analysis showed a slight tendency for the trials with the most extreme amplitude and highest costs to be most prone to accounting errors. trials is The biased slightly towards those effect is who Thus the final sample of eleven understood and performed best in the game. modest, however, and reinforces the conclusions drawn below regarding misperceptions of the feedback structure by the subjects. experiment and 4. In the 5. The parameters respectively. in the 6, as, p, and S' were estimated to the nearest The search was capturing the global simulations orders are restricted to the positive integers. minimum .1, .05, .05, and 1 carried out over a sufficiently large range to ensure of let^. units, i^- ^ J ^ 6. 1 Note that S' functions nonlinearity much Ot>0 means like a regression constant in eq. 32. However the residuals will not, in general, satisfy Let = and actual orders need not have a common an appropriate measure of goodness of fit. mean). The conventional R^ The alternative R^ = (estimated is r^ is used, the coefficient of correlation between estimated and actual orders (Judge et 7. Subjects were assigned positions randomly, and one-way ANOVA the not therefore where al. r is 1980). of the estimated parameters showed no strong relationship between the values of the estimated parameters and the position one plays significant at the 4.3% in the game (with the exception of as which was level). N.B. The long wave game runs on any IBM PC or compatible computer; disks are available from the author. The estimation and simulations for both experiments were carried out on Macintosh computers using the TrueBASIC language. The data and computer programs are available from the author upon request. To avoid plots shown in figure discarding the first 1500 weeks. The the first 1000 4 were generated by simulateing the model transients, the phase for 10,000 periods and 9000. Similarly, the phase plots shown in figure 8 were simulated for first 500 were discarded were discarded. for all except the Bassbeer factory, for which D-3921 25 References Bakken, B. 1987. Learning System Structure by Exploring Computer Games, unpublished manuscript, System Dynamics Group, Sloan School of Management, MIT, Cambridge MA 02139. Beylich, A. 1986. On the Modelling of Vehicular Traffic Flow. In C. Kilmister (ed.) Disequilibrium and Self-Organization. Dordrecht: D. Reidel. Cyert, R. and J. March. 1963. A Behavioral Theory of the Firm. Englewood Cliffs N.J.: Prentice Hall. Davis, H. L., S. J. Hoch, and E. K. Easton Ragsdale. 1986. An Anchoring and Adjustment Model of Spousal Predictions. Journal of Consumer Research. 13:25-37. Day, R. 1984. Disequilibrium Economic Dynamics: Journal of Economic Behavior and Organization. A Post-Schumpeterian Contribution. 5, 57-76. Day, R. 1982a. Irregular Growth Cycles. American Ecoru>mic Review. 72, 406-414. Day, R. 1982b. Complex Behavior Einhom, H. J., in System Dynamics Models. Dynamica. and R. M. Hogarth. 1985. Ambiguity and Uncertainty 8, 82-89. in Probabalistic Inference. Psychological Review. 92:433-461. Forrester, J. W. Forrester, J. W. and 1961. Industrial Dynamics. P. M. Senge. Models. TIMS Studies Frankel, J. in the Cambridge, MA: MIT Press. 1980. Tests for Building Confidence in System Management Sciences. Dynamics 14:201-228. A. and K. A. Froot. 1987. Using Survey Data to Test Standard Propositions Regarding Exchange Rate Expectations American Economic Review 77:133-153. Hines, J. 1986 A Behavioral Theory of Interest Rate Formation. Working paper Sloan School of Management, MIT. Hogarth, R. M. 1987. Judgement and Choice. 2nd ed. New York: Wiley. WP- 177 1-86, 26 D-3921 M. and M. W. Reder Hogarth, R. (eds). 1986. The Behavioral Foundations of Economic Theory Journal of Business. 59:S181-S505. Jarmain, W. Johnson, E. E. (ed.) 1963. Problems Dynamics. Cambridge, in Industrial MA: MIT Press. and D. A. Schkade. 1987. Heuristics and Bias in Utility Assessment. J. Unpublished manuscript, Wharton School, University of Pennsylvania, Philadelphia. Judge et al. 1980. The Theory and Practice of Econometrics. New York: Wiley. Lopes, L. L. 1981. Averaging Rules and Adjustment Processes: The Role of Averaging in Inference. Report 13, Wisconsin Human Information Processing Program, University of Wisconsin, Madison. Lucas, R.E. 1976. Econometric Policy Evaluation: (eds). A Critique. The Phillips Curve and Labor Markets. Supplement In K. Brunner and A. Meltzer to the Journal of Monetary Economics. May, R. M. 1976. Simple Mathematical Models with Very Complicated Dynamics. Nature. 261(5560):459-467. Bounded Omega. 11:131-142. Morecroft, J. 1983. System Dynamics: Portraying Morecroft, J. 1985. Rationality in the Analysis of Behavioral Simulation Models. Rationalit>'. Management Science 31:900-916. Mosekilde, E., D. Rasmussen, H. Jensen, in a Generic Sturis, and J. Jespersen. 1987. Chaotic Behavior Management Model. European Journal of Operations Research. Forthcoming. Mosekilde, E. and E. Larsen. Dynamics Review. Nicolis, G. J. and I. 1987. Chaotic Behavior in the Beer Distribution Game. System 4(1). Prigogine. 1977. Self-Organization in Nonequilibrium Systems. New York: Wiley. Plott, C. 1986. Laboratory Experiments in Economics: The Implications of Posted Price D-3921 27 Science. 232 (9 May), 732-738. Institutions. Prigogine, and M. Sanglier I. (eds.) 1987. Laws of Nature and Human Conduct. Brussels: Task Force of Research Information and Study on Science. Rasmussen, a Simple S., E. Mosekilde, and Model of the J. D. Sterman. 1985. Bifurcations and Chaotic Behavior Economic Long Wave. System Dynamics Review. 1(1):92-1 in 10. Richardson, G.P.and A.L. Pugh. 1981. Introduction to System Dynamic Modeling with DYNAMO. Cambridge MA: The MIT Ruelle, D. 1984. Strange Attractors. Adam Press. In P. Cvitanovic (ed.)Universality in Chaos. Bristol: Hilger, Ltd. Shaw, G.K. 1984. Rational Expectations. New Simon, H. A. 1979. Rational Decisionmaking York: in St. Martins Press. Business Organizations. American Economic 69:493-513. Revie^^. Simon., H. A. 1984. The Behavioral and Rational Foundations of Economic Dynamics. Journal of Economic Behavior and Organization. 5:35-55. Smith, V. 1982. Microeconomic Systems as an Experimental Science. American Economic Review. 72, 923-955. Sterman, J. D. 1984. Instructions for Running the Beer Distribution Game. System Dynamics Group working paper D-3679, Sloan School of Management, MIT. Sterman, J. D. 1985a. A Behavioral Model of the Economic Long Wave. Journal of Economic Behavior and Organization. 6:17-53. Sterman, J. D. 1985b. A Skeptic's Guide to Computer Models. Working Paper D-3665, System Dynamics Group, Sloan Scholl of Management, MIT, Cambridge Sterman, J. MA 02139. D. 1986. The Economic Long Wave: Theory and Evidence. System Dynamics Review. 2(2):87-125. 28 D-3921 Sterman, J. D. 1987a. Testing Behavioral Simulation Models by Direct Experiment. Management Sterman, J. Science. D. 1987b. Misperceptions of Feedback WP- 1899-87. paper Sterman, J. 33(12). in Dynamic Decisionmaking. Working Sloan School of Management, MIT, Cambridge MA. D. 1987c. Managerial Behavior in Dynamic Decisionmaking: Further Misperceptions of Feedback. System Dynamics Group working paper D-3919, Sloan School of Management, MIT. Sterman, J. D. 1987d. Expectation Formation in Behavioral Simulation Models. Behavioral Science. 32:190-211. Sterman, Game J. D. and D. L. Meadows. 1985. STRATEGEM-2: A Microcomputer Simulation of the Kondratiev Cycle. Simulation and Games. 16:174-202. Thompson, J. and H. Stewart. 1986. Nonlinear Dynamics and Chaos. Tversky, A. and D. Science . Kahneman 1974. Judgment New York: Wiley. Under Uncertainty: Heuristics and Biases. 185 (27 September): 1124-1 131. Zeeman, E. C. 1977. Catastrophe Theory: Selected Papers, 1972-1977. Reading, MA: Addision Wesley Publishing Co. Zimmerman, R. 1986. The Century. In C. Kilmister Transition from (ed.). Town to City: Metropolitan Behavior in the 19th Disequilibrium and Self-Organization. Dordrecht: D. Reidel. 29 D-3921 Figure 1. The generic stock-management system. u Exogenous Variables Stock Acquisition System Expected Loss Rate Ordering Heuristic D-3921 Figure 2. 30 The stock management system applied to aggregate capital investment. portrays a firm which produces capital plant and equipment. is exogenous, as multiplier is is the delivery delay for acquiring introduced, the firm's the delivery delay for capital new The demand capital stock. The model for the firm's product When the self-ordering orders for capital add to the exogenous demand, and the firm's own delivery delay. own becomes CAPITAL N ^ DISCARDS o CARTAL STOCK SUPPLY UNE CAPITAL ACQUISITIONS / S^ ill ( DELIVERY L DELAY 7 fpRooucnoNj \ J BACKLOG K-not _ -/OTHER \ORDERS D-3921 31 Figure 4 3. Typical results of the macroeconoinic experiment. iitKaMaMMMMMa < < • It II a 14 a 32 D-3921 Table 1. Macroeconomic experiment: Summary of results. 35 D-3921 Table Trial 1 Score 2. Macroeconomic experiment: Estimated parameters. as Std. Error* asL Std. Error* R2 Mode D-3921 34 Figure 4. Simulation of the decision rule with estimated parameters. HOOn Game 36 1200n Gome 18 1200 10001000 600 600 1200 800 1200 eoo 100 Production CapecUij Proaucllon Capacity * eoo Game 12 750 c TOO o 880 o a. ^ 600 •> o 550 500 1230 150 700 600 500 eoo Production Capacity 1200-1 Game 10 100 800 Production Capacity 1200 1250 Production Capacity 1270 35 D-3921 5. Modes of the system in parameter space. Lower graph zooms in to area between (0,1). Figure b- a U- J3 L. 1 : : 1? 37 D-3921 Figure 7a: 7. T^'pical experimental results of the Compare Beer distribution Game. week 5 customer orders rise from 4 to 8 cases per week. against the oscillations in the subjects' orders (7c). Customer Orders. In 3025 - ro15 ' J - 5 - - Figure 7b. Key a to experimental results (figure 7c). Factory ^ Distributor _ Wholesaler Retailer 5 10 U 20 25 30 Orders placed by sector. From bottom to top, R. W. D, F, each offset by 1 5 cases/week. Major tick-marks=l5 cases/week. Minor tick-marks=5 cases.'week Initial orders=4 cases/week in all sectors. :5 Weeks Factory Effective Inventory o c by seaor. Effective lnventory=lnventory-Backlog. 01 > c Distnbutor From bottom to top, R, W. D, F, each offset by 40 cases. o o Wholesaler Maior tick-marks=40 cases. Minor tick-marks=lO cases. Initial inventorysl2 cases in Retailer 3 10 13 20 23 Weeks 30 33 all sectors. 38 D-3921 Figure 7c. Typical experimental results. Effective Inventory Orders »- D-3921 39 Table 3. Summary of experimental results, Beer Distribution Data are averages of Total COSTS Mean Benchmark* Ratio 1 1 Game. trials. Retailer Wholesaler Distributor Factory D-3921 40 Table Trial & Position Bassbeer 6 4. Estimated parameters, Beer Distribution as p S' Game r2 Mode 41 D-3921 Figure 8. Simulation of decision rule with estimated paraemters, Beer Distribution Game. 50 60 HemekenJ Retailer Coors Factory 10 50 20 20 -20 10 -10 -60 -10 -eo -20 20 Factory Inventory Factory Inventory 80 1 — I 42 D-3921 Beer Distribution Game. Modes of Figure 9. the system in parameter space. 1.0 Stable 0.8 • Period • Period 5 • 1 Period 12 Chaos 0.6 0.4- 0.2 D B o f 0.0 0.0 0.2 » * — n I —D- o 1 0.4 0.6 a 0.8 1.0 _ 40 Stable • Period • Period 5 • Period 12 35 1 30 Cliaos 25S' 20- • B I . 15 10 5 — 0.0 0.2 0.4 a g I 0.6 0.8 Date Due iOV 151990 \»*:m^ o^ Lib-26-67 MIT LIBRARIES 3 TDfiD DD4 T2fi BEfi