all model structures consist of two parts: assumptions about the physical and institutional...
TRANSCRIPT
All model structures consist of two parts:
•Assumptions about the physical and institutional environment
•Assumptions about the decision processes of the agents
Includes the model boundary and stock and flow structures of people, material, money, information, and so forth that characterize the system
Forrester’s Urban Dynamics sought to understand why America’s large cities continued to decay despite massive amounts of aid and numerous renewal programs
Refer to the decision rules that determine the behavior of the actors in the system
In Urban Dynamics, these included decision rules governing migration and construction
Institutional structure of a system is relatively straightforward.
Subtle and challenging
To be useful simulation models must mimic the behavior of the real decision makers so that they respond appropriately, not only for conditions observed in the past but also for circumstances never yet encountered
Modelers must make a sharp distinction between decisions and decision rules
Decision rules are the policies and protocols specifying how the decision maker processes available information
Decisions are the outcome of this process
It is not sufficient to model a particular decision.
Modelers must detect and represent the guiding policy that yields the stream of decisions
Every rate in the stock and flow structure constitutes a decision point, and• The modeler must specify precisely the decision
rule determining the rate
Can be thought of as an information processing procedure
The inputs to the decision process are various types of information or cues• The cues are then interpreted by the
decision maker to yield the decision• Decision rules may not use all available
information
Cues used to revise prices in the department store case include wholesale costs, inventory turnover, and competitor prices
Department store pricing decisions do not depend on interest rates, required rates of return, store overhead, trade-offs of holding costs against the risk of stock-outs, estimates of the elasticity of demand, or any sophisticated strategic reasoning.
Is our model a descriptive model or is it a prescriptive one?
Recall:• Descriptive models…tell it like it actually is• Prescriptive models…tell is like it should be
Mental models of the decision makers
Organizational, political, personal, and other factors, influence the selection of cues from the set of available information
The cues (information) used is not necessarily processed optimally
All human behavior can be viewed as involving participants who maximize their utility from a stable set of preferences and accumulate an optimal amount of information
Not only do people make optimal decisions given the information they have, but they also invest exactly the optimal time and effort in the decision process, ceasing their deliberations when the expected gain to further effort equals the cost
The Baker Criterion: The inputs to all decision rules in models must be restricted to information actually available to the real decision makers
Senator Howard Baker: What did he (Nixon) know and when did he know it??
Must ask, “What did they know and when did they know it?”
To properly mimic the behavior of a real system, a model can use as an input to a decision only those sources of information actually available to and used by the decision makers in the real system
First, no one knows with certainty what the future will bring
Second, perceived and actual conditions often differ
Third, modelers cannot assume decision makers know with certainty the outcomes of contingencies they have never experienced
All variables and relationships should have real world counterparts and meaning
The units of measure in all equations must balance without the use of arbitrary scaling factors
Decision making should not be assumed to conform to any prior theory but should be investigated firsthand
Fractional Increase Rate
Fractional Decrease Rate
Adjustment to a goal
RI = g * S
Here, RI is an input rate, g is some fraction (<1) and S is the stock that accumulates RI
Examples Birth rate = birth rate normal * Population Interest Due = Interest Rate * Debt Outstanding
These examples all generate first-order, ________ loops.
By themselves, these rates create exponential growth
It’s never a good practice for these rates to be anything other than non-negative
RO = g * S
Here, RO is an output rate, g is some fraction (<1) and S is the stock that is depleted by RO
Examples Death rate = death rate normal * Population Death rate = Population / Average Lifetime
Left to themselves these rates generate exponential decay
Left to themselves, these rates create first-order, negative feedback loops
RI = Discrepancy / AT = (S* - S) / AT
Examples• Change in Price = (Competitors price –
Price) / Price Adjustment time• Net Hiring Rate = (Desired Labor – Labor) /
Hiring Delay• Bldg heat loss = (outside temp – inside
temp) / temp adjustment time
Generates exponential goal-seeking behavior
Is also considered a first-order, negative feedback loop
Often the actual state of the system is not known to decision makers who rely instead on perceptions or beliefs about the state of the system• In these cases, the gap is the difference between
the desired and the perceived state of the system
The Stock Management Structure: Rate=Normal Rate + adjustments
Flow = Resource * Productivity
Y = Y * Effect of X1 on Y * Effect of X2 on Y* … * Effect of Xn on Y
Rate = Normal Rate + Adjustments
If the input rate is RI = (S* - S) / AT, and the output rate is RO , then the steady state equilibrium will be S = S* - RO * AT
To prevent this the stock management structure adds the expected outflow to the stock adjustment to prevent the steady state error:
Inflow = Expected outflow + Adjustment for Stock
The flows affecting a stock frequently depend on resources other than the stock itself
The rate is determined by a resource and the productivity of that resource
Rate = Resource * Productivity, or Rate = Resource/Resources Required
per Unit Produced
Production = Labor Force * Average Productivity
These are called MULTIPLICATIVE EFFECTS
Examples:
Rate = Normal Fractional Rate * Stock * Effect of X1 on Rate * … * Effect of Xn on Rate
Birth Rate = Birth Rate Normal * Population * Effect of Material on Birth Rate * Effect of Pollution on Birth Rate * Effect of Crowding on Birth Rate * Effect of Food on Birth Rate
A reference year of 1970 was defined
Normal fractional birth rate was the world average in the reference year
All of the effects were normalized to their 1970 values, making those normalized values equal to 1
Create nonlinearities
Forrester really believes the effects are multiplicative
As an alternative consider additive effects:
Example: Change in wage = Fractional Change
in Wage * Wage Fractional Change in Wage = Change
in Wage from Labor Availability + Change in Wage from Inflation + change in Wage from Productivity + Change in Wage from Profitability + Change in Wage from Equity
Linear formulations are common because such formulations are simple
Multiplicative formulations are generally preferable and sometimes required
The actual relationship between births and food, crowding, or pollution is typically complex and nonlinear
Both are approximations to the underlying, true nonlinear function: Y = f(X1, X2, …, Xn)
Each approximation is centered on a particular operating point given by the reference point Y* = f(X1*, X2*, …, Xn*)
Will be reasonable in the neighborhood of the operating point but increasingly diverge from the true, underlying function as the system moves away from it
Additive assumes the effects of each input are strongly separable
Strong separability is clearly incorrect in extreme conditions
In the birth rate example, births must be zero when food per capita is zero no matter how favorable the other conditions are
The additive formulation can never capture this
Fuzzy MIN Function
Fuzzy MAX Function
Floating goals
A rate or auxiliary is determined by the most scarce of several resources
Production = MIN(Desired Production, Capacity)
Generally, Y = MIN(X, Y*), where Y* is the capacity of the process
The sharp discontinuity created by the MIN function is often unrealistic
Often the capacity constraint is approached gradually due to physical characteristics of the system
A fuzzy MIN function will accomplish this for us so that there is not sharp discontinuity
Analogous to fuzzy MIN function Hiring Rate = MAX(0, Desired Hiring
Rate) prevents Hiring Rate from ever gong negative
Useful in situations where decision makers want to keep a variable Y at its desired rate even as X falls to zero
The goal moves toward the actual state of the system while the actual state of the system moves toward the goal.
State of thesystem
Net Changein State
DesiredState ofSystem
Net Changein Desired
State
-
-
+
+
StateAdjustment
Time
GoalAdjustment
Time
B
StateAdjustment
R
Floating GoalSpiral
B
GoalAdjustment
Initial DesiredState of System
Initial State ofSystem
Performance and Goal
1,000
750
500
250
0
0 5 10 15 20 25 30 35 40 45 50Time (Week)
State of the system : inivsdes1 UnitsDesired State of System : inivsdes1 Units
Nonlinear Weighted Average
Modeling Search: Hill-Climbing Optimization
Resource Allocation
Decision makers must optimize a system but lack knowledge of the system structure that might help them identify the optimal operating point
Examples: A firm wants to • maximize profit• Minimize costs• Maximize the mix of labor and capital
Can do this in simulated real time using a variant of floating goals
The model adjusts the mix in the right direction, toward a desired state.
This is called hill-climbing
State ofSystem S
Change inState ofSystem
DesiredState S*
Effect of ExternalPressures onDesired State
-+
+
StateAdjustment
Time
B
StateAdjustment
+
R
Goal Rev ision
ExternalPressures X
Sensitivity toExternal
Pressures
Increase inExternalPressure
Decrease inExternalPressure
+
-
Time forDecrease in
Pressure
Time forIncrease in
Pressure
Converges to local optima
Must start it from a number of different points in the search space to ensure that a global optimum is found
But that is NOT WHAT IS GOING ON HERE—THE SIMPLE TECHNIQUE USED HERE IS JUST A VARIANT OF THE 1ST ORDER NEGATIVE FEEDBACK GOAL SEEKING STRUCTURE
Price
Change inPrice
IndicatedPrice
Effect ofDemand/Supply
Balance on Price
-+
+
PriceAdjustment
Time
B
PriceAdjustment
+
Demand
Supply
SupplyElasticity
DemandElasticityR
PriceDiscov ery
-
ReferenceDemand +
Demand/Supply Balance
+
+Sensitivity of Price to
Demand/SupplyBalance
ReferenceSupply
+
+
-
B
DemandResponse
B
SupplyResponse
Change inReferenceDemand
DemandCurve Shift
Time
InitialReferenceDemand
ReferencePrice -
<ReferencePrice>
-
Price
200
170
140
110
80
-2 -1 0 1 2 3 4 5 6 7 8 9 10Time (Period)
Price : price1 $/unit
Price vs. Indicated Price
200
170
140
110
80
-2 -1 0 1 2 3 4 5 6 7 8 9 10Time (Period)
Price : price1 $/unitIndicated Price : price1 $/unit
Demand vs. supply
125
116.25
107.5
98.75
90
-2 -1 0 1 2 3 4 5 6 7 8 9 10Time (Period)
Demand : price1 Units/PeriodSupply : price1 Units/Period
All outflows require First-Order Control
Avoid IF..THEN..ELSE Formulations
Disaggregate Net Flows
Real stocks such as inventories, personnel, cash and other resources cannot become negative
Outflow rates must be formulated so these stocks remain nonnegative even under extreme conditions
Do so requires all outflows to have first-order control
Means the outflows are governed by a first-order negative feedback loop that shuts down the flow as the stock drops to zero
Examples:• Outflow = MIN (Desired Outflow, Maximum
Outflow)• Outflow = Stock / Residence time• Maximum Outflow = Stock / Minimum Residence
time
Sterman doesn’t like these because they introduce sharp discontinuities into your models, discontinuities that are often inappropriate.
Individual decisions are rarely either/or
In many cases the decision is a compromise or weighted average of competing pressures
They create conditional statements that are often difficult to understand, especially when the conditions are complex or nested with others
DISCUSSION