System Dynamics: Systems Thinking and Modeling for a Complex World

Source: System Dynamics: Systems Thinking and Modeling for a Complex World, MIT OpenCourseWare, 55:57, uploaded 2021-11-17, Watch Later position 35.

James Paine opens this MIT workshop by separating two terms that often get folded together. Systems thinking is the broad idea that actions, decisions, and events sit inside webs of dependence. System dynamics is a set of methods for making those relations visible, testing them, and using the results to change a social system. The two overlap, and Paine spends the lecture moving from the first idea towards the second.

From engineering problems to social systems

Paine introduces himself as a second-year PhD student in MIT’s System Dynamics Group. Before returning to academia, he worked as a chemical engineer at GE-Hitachi nuclear energy, as an industrial engineer and Lean Six Sigma specialist at Inmar, and as a product manager for HanesBrands’ Maidenform and Playtex 18 Hour brands. His research in behavioural operations management looks at how people make decisions inside supply chains. The presence of people, he says, belongs inside the model because human decisions shape the system that those decisions try to control.

That background leads to his account of the field. Jay Forrester brought ideas from control theory and computing into business and social systems at MIT. In 1958, Forrester published Industrial Dynamics after studying problems that resisted a purely engineering solution. A project with General Electric began with a three-year boom-and-bust cycle in employment at an appliance plant. The plant repeatedly hired large numbers of workers and then saw mass resignations. Forrester hand-coded a model to ask where the pattern came from.

Forrester later applied the same approach to larger social systems through the WORLD2 simulation and his work with the Club of Rome. Paine connects that work to The Limits to Growth, which he describes as a major system dynamics book with a contested public reception. Its simulations were often read as concrete forecasts. Paine gives the book a different role: it explores how resource renewal, regeneration, and the larger structure of society can produce patterns of overshoot and collapse. The model describes possible modes of behaviour under particular assumptions. It does not settle the date of a future event.

Paine’s own examples show how wide the field can be. One project models the pressures faced by nonprofit organisations and argues that a performance system can make some goals structurally impossible for a nonprofit manager. Another studies inventory oscillation in the Beer Game and asks how algorithms might control supply chains that contain human ordering decisions and delays. That work uses Python, TensorFlow, and optimisation alongside system dynamics ideas. The method set can change with the problem.

Closing the open loop

Paine begins with a familiar decision process. Someone identifies a problem, gathers data, evaluates alternatives, chooses a solution, and implements it. The sequence assumes that a decision changes the state of a system and brings the desired result. Supply chains make the missing relations easy to see. A change to one process alters a supplier’s decisions, changes an order stream, and affects other processes that sit outside the original team. A goal such as 3.4 defects per million opportunities can disappear as soon as the organisation reaches it and chooses a new target.

The effects also arrive with delays. A choice made today can change the conditions that shape the same choice later. The process therefore returns to its starting point through feedback. Paine treats this loop as the basic move in systems-level thinking. A decision exists inside a system whose boundaries determine which effects the decision-maker sees.

He defines a system as a set of interdependent parts with a common purpose, then adds a practical qualification. A boundary has to be drawn. If the definition expands without a purpose, every atom in the universe becomes part of one system and the model loses its use. The purpose gives the modeller a reason to include some parts and leave others outside.

Social and economic systems make this boundary work difficult. They change over time, connect decisions across distance, respond through feedback, contain nonlinear relations, and operate with limited information and ambiguity. A simple diagram of water flowing from one stage to the next leaves out the return path that changes the first stage. Systems thinking supplies a way to close that loop.

The first step is to elicit mental models. Paine asks the modeller to treat each decision-maker as someone acting rationally with the information available at the time. The task is to understand why the person sees a choice as sensible, then place that decision inside the larger system and trace how its consequences return to the person. Simulation helps because it lets a group test changes to a social process without waiting years for the real system to respond.

Paine gives simulation a modest purpose. A model does not have to reproduce the whole physical universe to be useful. He spends time trying to make his models match reference data, then warns that accuracy and usefulness answer different questions. A useful model identifies the parts of a feedback structure where a policy can change the behaviour that matters. It can be wrong about many details and still show where a decision has leverage.

Structure, behaviour, and the attribution error

Paine repeats the phrase “structure generates behaviour”. The structure includes physical arrangements, the information available to people, and the mental models through which they turn inputs into decisions. He compares the people inside a system to small controllers that receive information and produce actions, while warning that the conversion between the two needs to be modelled rather than assumed.

This principle changes how a modeller interprets blame. The fundamental attribution error leads people to explain an event through the character of the person involved and to overlook the system that made the action seem reasonable. Paine uses driving as the example. When another driver cuts him off, his first response is to blame that driver. When he has to reach a preschool because his child is ill, he can imagine himself making the same decision and asking for forgiveness afterwards. The action has an underlying reason even when it frustrates someone else.

The System Dynamics Group writes its counter-assumption on a whiteboard: people are intelligent, capable, concerned with doing their best, acting with integrity, and willing to learn. Accepting that assumption shifts the modelling question towards the conditions under which people work. The system may need to change before the behaviour can change. Paine includes the caveat that the group still has to remind itself of this principle and does not apply it perfectly.

He then places three layers beneath visible events. Events are what a person sees in the moment. Repeated events form patterns of behaviour. The structure beneath those patterns produces them. A series of headlines about a drunk trader, OPEC rumours, a tanker attack, or economic worry can explain individual movements in oil prices. A longer view shows an up-and-down boom-and-bust pattern, with the noise around that pattern increasing over time. The deeper question concerns the physical structure, information, and mental models that generate the pattern.

Learning in such an environment is hard because people act with incomplete information and time delays. A person can only choose from the information that reaches them through the system. Expecting a better decision without changing that information structure asks for a kind of omniscience. Systems thinking therefore gives rationality a context rather than treating it as a fixed personal trait.

The spiral of model building

Paine presents system modelling as an iterative spiral. A modeller begins with a reference mode, a behaviour that can be observed and described: a particular decision under particular inputs producing a particular output. The first model tries to reproduce that behaviour. A second reference mode may reveal that the initial structure leaves something out, which sends the modeller back to add structure and test the model again.

Interviews make the need for this process concrete. In Paine’s nonprofit research, six stories may describe three outcomes across two possible modes of behaviour. He advises choosing one mode, modelling it, then adding the next. The modeller can compare the structures afterwards and decide how they belong together. Sensitivity analysis then becomes a way to update the modeller’s own assumptions about the problem.

This is where stock-and-flow diagrams enter the lecture. Paine stresses that system dynamics is a field of thinking and practice, while stock-and-flow diagrams are one common modelling choice. They are useful because they make accumulation, feedback, and delay explicit. They are not the whole field.

Causal links and loops

A causal link says that a change in one variable changes another. Paine starts with production, inventory, and shipments. More production raises inventory, while more shipments lower it. More salespeople can raise the number of booked orders, and a higher price can lower it. Births add to a population and deaths reduce it.

Ambiguity exposes missing structure. A larger sales force may raise orders until salespeople compete for the same customers, overlap in their work, or irritate customers with too many calls. Once the sign of a link can change, the diagram needs a mechanism between the two variables. The same problem appears in the familiar relationship between ice-cream sales and murder rates. The two may rise together, yet average temperature can drive both: warmer weather increases outdoor interaction and ice-cream consumption.

Paine takes questions from the room about the word “causal”. His working definition is deliberately loose for social-system modelling. A link counts as causal enough for the model when a change in one direction produces a consistent change in the other, all else being equal. A physicist in the audience points out that causality depends on the assumptions that define a system. Paine agrees and says that his own distinction between causality and correlation is a practical convenience rather than a final account of the universe. When a link changes sign under different conditions, the modeller has to split it into more specific structure.

The lecture then assembles a loop around employee skill, customer satisfaction, complaints, and a manager’s time. Higher employee skill raises customer satisfaction, which lowers complaints and frees manager time for coaching. More coaching raises employee skill again. A small change can therefore travel around the loop and return with a larger effect. Paine calls this a reinforcing loop.

A balancing loop acts towards a goal or a previous state. Market attractiveness can bring in more competitors, which can lower product prices and reduce profits. A goal-seeking loop compares actual performance with a desired level and acts on the gap. To classify a loop, Paine suggests tracing a small increase or decrease around the full circuit. An even number of negative links produces a reinforcing loop; an odd number produces a balancing loop. The result depends on getting the link signs right.

Stocks, flows, and memory

A stock is a quantity that accumulates and therefore gives a system memory. Paine draws it as a bucket with an inflow, an accumulation, and an outflow. The number of employees and the number of units in inventory persist across time. A bathtub makes the constraint clear: material in the middle can leave only through the outflow, or the inflow has to decrease.

Greenhouse gases in the atmosphere provide the policy example. The stock can fall through greater net removal or lower net emissions. The same distinction separates a balance sheet from a cash-flow statement, wealth from income and expenditure per unit of time, and vehicle production from the number of vehicles. The category depends on how the modeller defines the variable.

Interest rate is Paine’s deliberately tricky case. The word “rate” suggests a flow, yet he treats an interest rate as a price on money. It describes what a sum of money returns over a period and has the persistence of a stock in the model. The example shows why a label alone cannot settle the role of a variable.

The question of usefulness follows from these tools. Paine says system dynamics can model modes of behaviour, such as an oscillation whose amplitude and period change under different inputs. It has less value as a precise point predictor far into the future because small differences between the model and the world grow with distance from the present. The useful question concerns which structures affect the mode and which policies can change it. The model earns its place through policy insight rather than a promise of exact dates and values.

Fishbanks, the Beer Game, and the structure of experience

The workshop’s management flight simulator is Fishbanks, a game in which groups run deep-sea and coastal fishing businesses. Each group receives a small fleet, catches fish, and tries to make money. The description says that the management simulation was removed from the upload because of copyright restrictions. The available cut preserves Paine’s setup and a discussion after the exercise, which means the gameplay itself cannot be reconstructed from this video.

That discussion still carries the lesson. A participant observes that one team’s change happened too quickly because the group lacked a clear view of the equations behind the model. Paine says the information is imperfect yet useful, and suggests that a second round could give players more information under changed parameters. He connects the point to the Beer Game, a multi-echelon supply-chain exercise in which even experienced players reproduce the same outcome. MIT once ran a Beer Game championship with professors who taught the exercise, and prior experience did not free them from its result.

The underlying structure remained in place. Information can help people, yet information alone does not dissolve a feedback structure that keeps producing the same behaviour. Paine uses the Vasa as another example. The ship was designed by skilled builders, then sank after late additions enlarged the captain’s quarters, added sails, and added cannons. The additions changed the ship’s stability. The result was an unintended effect that belonged to the design process itself.

Learning the field and choosing a tool

Paine points viewers towards MIT courses 15.871, Introduction to System Dynamics, 15.872, which applies the ideas to more realistic situations and modelling tools, and 15.873, which places greater weight on business and policy. He recommends John Sterman’s Business Dynamics for its first two chapters and Peter Senge’s The Fifth Discipline Fieldbook for a broader introduction to systems thinking. The Limits to Growth remains useful as a historical case, provided its simulations are read as descriptions of possible behaviour under a structure rather than as dated prophecies.

He also names the paper “System Dynamics at Sixty: The Path Forward” as a current account of the field. Its point, in Paine’s presentation, is that system dynamics is a way of thinking whose software and communication methods can change with the work. A causal diagram can live in a specialist tool, Excel, or on a whiteboard. The first move is to write the quantity of concern in the middle of the board, ask what would make it rise or fall, add signed arrows, and follow the loops.

The resources include the Creative Learning Exchange, Tom Fiddaman’s MetaSD site, the System Dynamics Society, Vensim, Stella Architect, and NetLogo. Vensim and Stella Architect support stock-and-flow modelling. NetLogo offers an agent-based route for modelling many individual entities and observing what happens when they occupy the same environment. Paine’s rule for choosing among them is practical: use the tool that lets the model answer the question.

The final example is En-ROADS, a climate policy simulator developed by Climate Interactive and MIT Sloan. A large climate model runs behind its controls, yet its purpose is policy exploration rather than a point forecast. Users can change assumptions, inspect the temperature path, and ask which choices alter the outcome. Paine says the simulator has been used in conversations with government officials to elicit mental models and create a shared basis for climate action. The workshop ends with that same movement from a visible outcome to the feedback structure that could change it.

Further reading / references

• MIT OpenCourseWare course page for the workshop, slides, and course materials.
• John Sterman’s Business Dynamics, the textbook Paine recommends for the field’s basic concepts and modelling process.
• Peter Senge’s The Fifth Discipline Fieldbook, which Paine recommends as a broader systems-thinking guide.
• The Club of Rome’s The Limits to Growth, the book Paine connects to Forrester’s WORLD2 work and the history of system dynamics.
• John D. Sterman’s “System Dynamics at Sixty: The Path Forward”, named in the lecture as a recent account of the field.
• Fishbanks: A Renewable Resource Management Simulation, the management simulation whose gameplay is absent from this upload.
• Creative Learning Exchange, MetaSD, and the System Dynamics Society, the educational, modelling, and community resources Paine names.
• En-ROADS, the climate policy simulator shown at the end of the workshop.