Webinar Q&A | Applied System Dynamics for Students and Beginners
We had an insightful Webinar with the participation of Andries Botha and Nalini Pillay who guided us in the first steps to learning Systems Dynamics.
You can watch the webinar recording and download the presentation here.
Here are the answers to questions asked live during the Webinar.
Learn more about the Seminar Series.
Q&A Webinar Applied System Dynamics for Students and Beginners
Answers by Andries Botha (AB) and Nalini Pillay (NP)
- How can you professionalize in System Dynamics without a formal education in the field, for example, a Master’s in System Dynamics or a Ph.D.?
AB: System dynamics is a tool to understand and solve problems. In itself, the language of stocks and flow is very simplistic. The underlying mathematics is also relatively simple – first-order differential equations for each stock that when combined results in higher-order differential equations. This means that a Master’s or Ph.D. will be difficult to establish. I would rather propose that the highly experienced members support new entrants to the field who wish to apply the methodology in solving the particular problem to which they plan to apply system dynamics. Personally, I use many tools to solve problems. System Dynamics has a place where decisions have to be made in an environment where uncertainty, complex interactions, feedback, and the time domain are important. Optimization will give you a solution set, but system dynamics will allow you to explore the domain and come up with a richer understanding of the problem that may result in the problem disappearing. I have no formal SD training but used it in both my MBA and Ph.D. work. Of course, efforts by the Society to formalize the underlying computational language constructs are a strong step to ensuring that the discipline is applied consistently.
NP: It is entirely possible to develop system dynamics modeling skills without a formal education, and making use of free course material such as on the MIT website, several books some of which are free, various websites, etc. Formal degrees largely provide recognition in the field once you have established skills through all the available resources, and enable credibility on international platforms. Even at the end of an MSc or Ph.D., a system dynamicist will never reach a point of saturation, SD is a continuous journey of learning. Finding a mentor or coach is important and belonging to communities where engagement and discussions are possible, also assists. There are a few tips and tricks that can be learned in these discussions.
- What are the main challenges to model Public Policies in developing countries where quantitative data is scarce and some qualitative data is available?
AB: The purpose of data should never be to populate the time domain in system dynamics. Understanding the underlying structures and how people make decisions are more important. As such data is useful to estimate fixed conditions and should never replace decision nodes. Data is critical for calibration and this is where you need a single time-based data stream as input to compare your model against the behavior of an explicit output or set of outputs. The amount of output/result data is significant. The World Bank has a huge dataset that can be used for free. The first focus, understand the problem, formulate a model not on data but on the structure that drives decisions and behavior. Then use the minimum input data to generate output data. Compare the output data.
- How to study problems in which there is no documented data? For example problems of low resource and small communities? Only causal loop diagrams?
AB: Small communities have lots of data sources. Ask the people. The focus should never be exclusively on data. As an example. A school wanted to know what the implications would be if they double the size of the school. Finance was a factor, but not the only factor. We ended up building a simulation to demonstrate the implications of the proposed change to the ethos of the school. The results of this were so insightful that the school changed its curriculum to support the ethos and scrapped the growth plans. I still owe them a financial model – the only hard data they had.
NP: The way in which we have dealt with projects that do not have adequate data, is by finding proxies that can be used. It is true that you can build a system dynamics model such that the structure can generate system behavior specific to the system problem even without the empirical data, however, it is necessary to have a quantitative data fitting, with approximations and assumptions if necessary. An example would be trying to determine the relationship between income deciles and residential electricity consumption. We may have data on household appliance electricity consumption but to understand the dynamics of electricity consumption of people on various income levels means that we understand behavior, not necessarily data. We know that once income levels increase, the middle-class population will start buying goods such as air conditioners and more electricity-dependent goods, the lower-income levels may move from biofuel energy sources to stoves, etc. This means that we can then relate the income levels to the appliances that would most likely be used due to expected behavior and be able to determine a relationship between income levels and residential electricity consumption.
- How do we know that the model needs no more variables? Is it always good to have as many variables as possible? Is it up to the modeler to choose the variables that matter the most?
AB: How did Van Gogh know the picture was complete? I am a strong believer in small models with simple structures that guide. Yes, there is always the temptation of adding one more variable. The minimum number that makes to model run and allows you to answer the particular question you are asking is the correct number. For example, most population models are fine with one inflow, one stock, one outflow, and two converters. If you want to add immigration and emigration you need to add two more stocks and two more convertors. Building a model in a country where the rate of immigration and emigration is high will require you to do so. If not, rather move that to outside the boundaries of the model you are building. Establishing the boundaries of a model is of extreme importance.
NP: Level of understanding (y-axis) is a function of Complexity (X-axis). As complexity increases the level of understanding increases but you reach the peak of the bell curve after which your level of understanding actually starts decreasing because the model is now too complex. So having as many variables as possible is not necessarily a good thing. The selection of variables should be determined based on workgroup engagements (involving a multidisciplinary team), causal loop diagrams, the SAM, and the focusing question as well as the scope of the project. Tabulating the endogenous, exogenous, and significant other (but excluded) variables may also assist in establishing a system boundary and ensure that you do not have scope creep and have included the relevant variables in the model framework. Of course, the number of variables can still then increase or decrease within 20% of the initial number of variables identified.
- Have you ever had a third party evaluate your model and modeling process?
AB: Yes. It is problematic unless you clearly stipulate the scope and the reviewer understands both the problem and the discipline. In one case I built a model with my mentor. We simulated the workings of a consulting firm. We ended up with two stock-flow streams that simulated elapsed time. I focused on time as a resource – hours to complete projects. He focused on hours as a systematic delay. Clients would request a proposal, but the sale would be delayed for up to 3 months. My resource time would continue to be available while we waited for the project to start. The model that combined the two perspectives gave us a good reflection of reality. When we handed the model over to a decision support specialist, the two-time flows were questioned and removed, with the result that the model became useless. In another case, the reviewer did not understand the nature of solving differential equations empirically and stated that the model does not work when dt>1. Of course, it would not. The smaller the dt the more accurate the solution. In this case, there was a threshold of dt=0.125. So yes, but be very careful of understanding the problem and domain being modeled and make sure the reviewer also understands the fundamentals really well.
NP: We have built several system dynamics models, the results of which have been explained in publications and papers. Once shared, there is generally a lot of interest from third parties to rigor check the model and process. One such example is a system dynamics model of a Hydro pumped storage power station where we were investigating the potential to use the scheme as baseload by simulating additional units. When we presented it at one of the society conferences, Andrew Ford (who developed the ELECTRIC1 model) was busy with a team to look at similar work in the U.S., he then asked if he could look at the model structure and interrogate the model and based on his analysis, wanted to collaborate on further work. Other models that we have built such as the Electric Vehicle Simulator was evaluated by EPRI as well as several research institutes with experience in the automotive industry. Our economic sector modeling process was evaluated by the Department of Energy in South Africa. If the modeling process is comprehensive and the model is robust, it will definitely withstand the scrutiny of any third party!
- Could you please elaborate a little bit on how you turn SAM into CLD (Causal Loop Diagram)? Thanks!
NP: A SAM shows the upstream and downstream variables and provides a high-level birdseye visual map related to the system problem. The Causal Loop Diagram shows the cause-effect relationships between variables. The SAM and the CLD are part of the suite of stakeholder engagement tools. However, if you would like to translate the SAM into a CLD, each box in the SAM is almost like a sub-model theme of what needs to be structurally modeled. So if I take a SAM box that refers to diesel tanks, in my CLD, I will find all the factors that affect the diesel tanks and illustrate the causality of those variables.
- Is it possible to obtain a copy of Nalini’s slides?
You can download it here
- How to deal with systems that are continually changing?
AB: The question here is what do you consider as a continually changing system? Sometimes we confuse changes in outputs to be a result of the system changing, but the underlying structure and decision rules may be stable, it is a perceived response or combined system complexity. For example, if I have a company that serves 5 destinations for one client. Allocation is an optimization problem. If I add return loads that become available, but may have to be waited for, I add a dimension of time, meaning a System Dynamics model will give a better solution. The location and activity of each truck will continually change, but the model covers that already. A second example, the price of oil follows a “random path”. Yet the underlying structure is fixed. The is a fixed amount of oil on the planet. We have explored a certain amount and a certain amount is economically viable for exploitation. If demand changes and we start exceeding the amount that is economically viable, the price goes up and people start exploring the next more expensive opportunities and start exploiting the opportunities that become economically viable. Now there is too much oil for demand and the price goes down. Lots of change in the oil price but no change to the structure. If you think what you are trying to model is continually changing, maybe you are not looking at the structural layer.
- Could you briefly explain why we should not interpret DT relative to time frames? (DT is the step size) E.g. I always think DT 1/12 is equivalent to a monthly timestep, is this wrong?
AB: My favorite subject – dt. There is a whole section on this in the book. Firstly, ignore the default settings. All time-domain simulation tools simulate in simulation cycle units. We assign days, hours, years, etc. arbitrarily. Decide what time units the problem you are working with will be most effective. For example, if you want to simulate the impact of the gap between a factory that works 5 days a week and a retail operation that works 6 or 7 days a week the obvious choice is to work in days, not weeks. Select 365 days to give you one year and off you go. Dt focuses on the solving of the differential equations in the stocks. You are trying to integrate non-linear curves with no obvious analytical solution. The way to do that is to take the first-time unit and create an estimation of the surface below. If the line is linear easy. If not, the estimate will have some error. The smaller the dt, the more accurate the surface estimation becomes. dt is just the method to make models more accurate. In the old days, the solution engines had constraints so you were limited. I have recently run a Lorentz butterfly model with 1 million data points with no problem. For an accurate solution of the differential equations, I used dt=0.0001. Awesome result.
- Is SAM the same as the sector map (Sterman)?
NP: I am not familiar with Stermans’ sector map. A SAM does not display cause and effect relationships or directional quantities linked to the variables but includes important upstream and downstream key variables, driving forces, and externalities specific to that environment.
- How can we tackle the unreliability of data before a certain time? By removing them I am only left with few years of data
AB: See above (Question 9). Structure of the problem over data. Always get the structure right. Don’t let data drive the model. Use data for calibration. All models are wrong, some are more wrong than others, but some are useful. I have a model that accurately forecasted annual global economic growth from 2000 to 2015, initialized with estimates of the global economy in 1800.
- Is there a way to let the computer improves the model structure (create and remove elements and connections) so it fits better the training data?
AB: None of the available System Dynamics software allows for the model to build itself. I spend more time with people who understand the system to define a structure that will provide insight. The algorithms for decision-making coming from these are in my opinion more valuable than any neural net can achieve. Case in point, in 1995 I proved a decision algorithm for a company. In 2019 they tried to replace it with the latest technology. After a huge investment, lots of calibration, and testing the new solution lasted for 3 months in the field and was then replaced with the original algorithm.
- Do you think the speed of adoption of System Dynamics methodology by the market is slow? If so, why is that?
AB: System Dynamics is easy but not simple. Or, System Dynamics is simple but not easy. I always tell people who attend my courses that I may do it as if it is obvious. It is not. Once you finish the course it will take you an hour to build a complex model that will never run. I am very good at it, so I can do it in 20 minutes. It will however take me 24 hours to think through the problem, another 8 or so to formulate the concept and then I will build a small start of the model in 10 minutes that will run. I will then expand. In my honest opinion, some people are hardwired to see the model in their heads, others are not. In my training, 4 out of 8 will never try, 3 will be able to run models and 1 out of 8 will be able to make the simulation sing.
NP: The speed of adoption is certainly slower in some countries compared to others, and there may be several factors impacting the rate of adoption. One factor is the number of formal education opportunities and courses accessible to interested parties in their countries. Another is the availability and support from System Dynamics communities. But the most important is the individual’s capacity to be able to understand the system problem and driving forces relevant to the environment; while being able to master the mechanistic aspects of the modeling software underscored by applied systems thinking philosophy. Our experience has shown that an engineering degree does not mean that a person will automatically be able to become a successful system dynamicist.
- Can you recommend a method/ technique for developing causal loops out of qualitative data?
AB: Listen, write the user/expert understanding out longhand. I usually do a two-hour interview with the user talking and talking. When you have all this information, draw the first few concepts. Every time you add a new perspective using a new color. I used china markers. Once they look pretty expand. If it stops looking pretty start on a new sheet. Always keep the past perspective. I use A3 for initial musings and A0 and bigger as my understanding grows. I have no formal method other than that.
NP: There is an unusual yet effective way to learn and practice Causal Loop Diagrams even without an initial team. The first is to take a narrative paragraph describing the system problem/related dynamics and then to highlight the keywords. Then write each term/keyword on a sticky note and start placing it on a blank piece of paper. Move the sticky notes around as you identify cause-effect relationships, then add additional variables to complete the gaps to close the feedback loops!
- How can we simulate a model which contains variables with very different units, including population (units people), vehicles
AB: Yes, another pet interest. When you start developing co-flow models the fun starts. Please make sure that only one element flow in each stock-flow train. For example, when you do pollution in a water system. The water flow/hydraulics is the main stock/flow. The mass balance of the chemicals is a separate but linked flow/flow. The link will be concentration – mass/volume. The water flow is volume/time unit etc.
- What is the title of the book you mention? Where can it be obtained? System Dynamics Society? Amazon?
AB: Applied System Dynamics With South African Case Studies
NP: The book is not yet available on Amazon or eBook stores but if you email firstname.lastname@example.org, we can provide a quote that will include courier costs to your location.
Honoring Excellence: A Glimpse into the Awards of the International System Dynamics Conference The International System Dynamics Conference brings together experts, practitioners, and students to exchange ideas, showcase real-world applications, and celebrate...
From Bergen to Global: UiB’s System Dynamics Group The System Dynamics Group, an autonomous research group at the University of Bergen (UiB) was established in 1971 by professor emeritus Svein Nordbotten. Inspired by the work of Jay W. Forrester, Nordbotten...
Consideo’s Insight into the Global Energy Transition: A Systems Perspective Complexities of Energy TransitionThe global energy transition is a complex challenge. The development of technology, the need for resources, the depletion of high-grade materials and...
Understanding and addressing the impact of structural racism on health is essential to building health equity. The October 2023 issue of Health Affairs, “Tackling Structural Racism in Health,” builds on the groundbreaking work contained in our February...
Recent Business cases
Fast-Track Cities Uses System Dynamics to Enhance HIV Care EXECUTIVE Summary Low levels of viral suppression at 69% for people with HIV make it hard to believe the 95% target level will be achieved by 2030 in St. Louis, USA. As a solution, Fast-Track Cities-STL opted...
System Dynamics Unravels ICU Tensions at the Portuguese Oncology Institute EXECUTIVE Summary The Portuguese Oncology Institute (IPO) faced a critical challenge in its intensive care unit (ICU) where doctors and nurses experienced high turnover due to tensions with...
System Dynamics Helps Farmers Escape Poverty Trap in Guatemala EXECUTIVE Summary Guatemala holds the 4th highest global ranking for chronic malnutrition, and climate change is intensifying the challenges subsistence farmers face in providing food for their families....