Special issue on advances in system dynamics modelling from the perspective of other simulation methods

1. Introduction and description of papers in the Special Issue

In 2017, System Dynamics celebrated the 60th anniversary of its founding by Jay Forrester, who sadly passed away at the end of 2016. One of the articles of the special issue (SI) is an account of his life as modeller and simulator. See Lane and Sterman in this issue. Since it was founded System Dynamics (SD) has grown over time to become an independent simulation methodology with applications in multiple areas and different levels of analysis but with a strong focus on strategic issues and policy formulation. SD models are essentially continuous-time simulation models driven by circular causality. The process for arriving at such representations helps to distinguish the field from other simulation approaches. For example, Wolstenholme (1990) defines System Dynamics as a “rigorous method for qualitative description, exploration and analysis of complex systems in terms of their processes, information, organisational boundaries and strategies; which facilitates quantitative simulation modelling and analysis for the design of system structure and control” (Wolstenholme, 1990, p. 3). Sterman (2000) suggests that SD modelling consists of discovering and representing feedback processes, which, together with stock and flow structures, time delays, and nonlinearities, determine the dynamics of a system.

The editors of this special issue have a long association with SD though each of us came to the subject in different ways. However, we all share a desire for SD to become an active member of the community of simulation methods such as DES or ABS. This desire is part of the natural evolution of any discipline to find common ground while appreciating and respecting the viewpoints of other disciplines. We were therefore grateful for the opportunity to collaborate on a special issue about SD here in the Journal of Simulation. The articles we received fall in two broad categories: reviews of multi-method approaches and demonstrations of multi-method approaches.

2. Findings from the reviews of multi-method approaches

There are three papers addressing reviews of SD combined with a variety of other methods, not only simulation. They span a range of perspectives from broad and general in Zolfagharian et al. to an SD perspective (with emphasis on areas of application) in Kunc et al., to a specific field of application, in Iturriza et al.

Zolfagharian et al. present a framework asking why, when and how to combine system dynamics with other methods. The authors draw on a systematic literature review (675 articles) to evaluate how different researchers have combined SD with at least one other method. They found that combining SD with another method enriched both the conceptualisation process and the simulation process. The review shows that the use of SD combined with other methods has grown substantially in recent years. The framework proposed considers four problem characteristics motivating the combination of SD with other methods: nature of the problem; variables and their causal relationships; diversity of agents involved; and context of the problem.

Kunc et al. conduct a computational literature review of System Dynamics. After selecting more than 700 papers from traditional simulation and OR journals, they clustered the papers into 51 topics studied in SD. Clearly, there are strong developments in healthcare and supply chain with important research in energy and economic systems. Over time, the SD field has more papers integrating SD with other methods, e.g., hybrid models, but there are not sustained contributions in areas such as statistical evaluation of parameters or automatic calibration. .

Iturriza et al. offer a survey of modelling methodologies, including SD, for analysing critical infrastructures. SD has a long tradition of focusing on macro-level problems, e.g., urban and world dynamics or epidemics. Critical Infrastructures (CIs) play a relevant role in both society and industry since they provide basic goods and services. What makes CIs important for SD modelling? It is their interdependent nature whereby a failure in one CI may spread rapidly to other dependent CIs resulting in a cascading effect that can generate high-impact crises. The paper surveys a variety of modelling methods with applications in financial services, telecommunications, transportation, emergency and security services as well as healthcare. SD and agent-based models were the two most commonly used modelling methods.

3. Demonstrations of how multi-method approaches feature in SD studies.

In this set of papers, we look at specific examples of SD used in combination with other methods. The examples include SD and DES in queuing models, SD and control theory, and optimisation methods for SD models.

Barlas and Özgün present a comparison of the use of SD and Discrete Event Simulation (DES) in queuing systems with feedback processes. This paper complements previous work comparing SD and DES, e.g., Morecroft and Robinson (2014). Barlas and Özgün conclude that DES provides better results for systems containing a small number of entities and servers but DES and SD provide similar outputs when the scale of the system increases (many servers and entities).

Schoenenberger and Tanase present a study combining SD with network controllability in order to identify effective leverage points for the design of influential policies. Network controllability is a concept from nonlinear dynamics and control theory that indicates the ability of a given set of feedback loops to steer a dynamical system towards a desired state. The basic process is to transform the structure of a System Dynamics model into a weighted network so variables and causal networks become vertices connected by edges. The method is illustrated using Forrester’s well-known World Dynamics model.

Linnéusson et al. focus on the links between the strategic and operational aspects of manufacturing equipment maintenance. The paper presents a model in which SD is combined with multi-objective optimisation to develop a decision support tool for production managers. The model simultaneously optimises three potentially conflicting objectives: maximising machine availability, minimising (short-term) maintenance costs, and minimising (long-term) consequential or knock-on costs. Conflicting short- and long-term objectives of this nature arise in many other contexts.

Parra et al. present the results from a study using metaheuristic optimisation methods to calibrate SD models. In this paper, the authors perform a comparison of four optimising heuristics (genetic algorithm, simulated annealing, Powell’s algorithm and a hybrid algorithm) with two different SD models. They conclude there is no best metaheuristic since the characteristics of the model have an important impact on their use. A useful contribution is the definition of criteria to evaluate metaheuristics such as objective function, regression coefficients, Theil coefficients and maximum errors. However, more research is needed to generalise the implications of using optimising heuristics with SD models.

4. The future of SD and mixed methods

The opening paper in this special issue tells the story of “The Model Simulator”, Jay Forrester. From his pioneering life of discovery and invention it is perhaps not surprising that a bold and distinct new approach to modelling and simulation arose. In looking to the future of system dynamics and its evolving relationship to other methods in the social sciences it is good to recall some of the field’s unmistakable and enduring features. Traditional system dynamics involves a distinctive “style” of modelling and analysis. It lays strong emphasis on clear visualisation and documentation of real-world feedback structure backed-up by rigorous yet easy-to-read equation formulations. Understanding of dynamics comes from careful narrative interpretation of simulations. From this foundation of rigour-with-accessibility it is then possible for SD-trained modellers to reach out securely to complementary methods and ideas, whatever form they might take. For example, mention has been made of connections to contemporary behavioural and resource-based views of the firm. These ideas, from modern economics and strategy, fit neatly with asset stock accumulation and the information feedback view of the firm found in traditional system dynamics (Forrester, 1961; Morecroft, 2015; Sterman, 2000). In addition there are references to complementary methods for the analysis of dynamic models, some originating in control theory and others in contemporary analytical and computationally intensive methods.

Beyond the construction and analysis of models lies their use by and for policymakers. Policy design is an important topic in system dynamics. It is part of the quest in the social sciences not only to understand phenomena in business and society but also to make things better. The established and traditional approach in system dynamics is pragmatic and behavioural. One could say it promotes a “nudge philosophy” of performance improvement. The essential idea is to test incremental change to existing feedback structure that improves (rather than optimises) system performance. There is practical merit in such a “satisficing” (or nudge) approach to policy design. It looks for performance improvement relative to a base case simulation. It acknowledges and accepts the enduring feedback structure that underlies the dynamics of the base case and seeks at least one high leverage point in the system where a small and understandable change in an operating parameter yields big improvements in performance, while leaving the vast majority of the already-established policy structure intact. Policy design can also be conducted with optimisation methods that are often sophisticated and quite technically demanding. Some of these methods are mentioned in this JoS special issue and others can be found in Rahmandad, Oliva, & Osgood (2015). Optimal policies normally require much more coordinating information than satisficing “nudge” policies in order to achieve performance improvements. The resulting tight control of interlocking operations presents significant implementation challenges that deserve careful consideration in technically optimal policy design.

There are important differences in worldview among different modelling and simulation methods. In reaching out for new and complementary approaches these “implicit worldviews of the modeller” need to be examined carefully. Typically they go beyond mere technical differences. For example, in comparing system dynamics and discrete-event simulation it soon becomes evident that they handle complexity in fundamentally different ways. System dynamics deals with “deterministic complexity” whereby the unfolding future is significantly pre-determined by enduring feedback structure. Discrete-event simulation deals with “stochastic complexity” whereby the unfolding future is partly and significantly a matter of chance arising from multiple interacting random processes. As modellers from different backgrounds reach out to each other, and seek to creatively combine their methods, it is important to understand the implications of contrasting and taken-as-given perspectives. The characteristic scope, scale and detail of factors included in a model, as well as its unfolding dynamics, arise naturally from the modellers’ chosen worldview (Brailsford, Churilov, & Dangerfield, 2014).

Martin Kunc
Warwick Business School, Coventry, UK
[email protected] John D. W. Morecroft
London Business School, London, UK
[email protected] Sally Brailsford
Southampton Business School, Southampton, UK
[email protected]