Simulating the behaviour of complex systems: computational modelling in ergonomics

Introduction

The complexity and scale of problems being tackled by ergonomists is growing. Work and societal systems are becoming increasingly reliant on technology, and the technologies themselves are becoming progressively sophisticated (Dekker 2011; Hancock 2019; Leveson 2004). On top of this, we are seeing a significant broadening of scope in terms of application areas, with ergonomics being proposed as a viable solution to large-scale societal and global scale issues (Salmon et al. 2019, Thatcher et al. 2018). As a result, systems ergonomics is receiving increasing attention, both within and outside of our discipline (Hulme et al. 2019; Karsh, Waterson, and Holden 2014; Salmon et al. 2017; Walker et al. 2017).

In line with this increasing interest, a core set of systems ergonomics methods are being applied in a diverse set of domains. Methods such as Cognitive Work Analysis (CWA; Vicente, 1999), the Event Analysis of Systemic Teamwork (EAST; Stanton, Salmon, and Walker 2018), Systems Theoretic Accident Model and Processes (STAMP; Leveson 2004), and the Functional Resonance Analysis Method (FRAM; Hollnagel 2012) are increasingly being applied to describe, evaluate, design and re-design sociotechnical systems to support human wellbeing and overall system performance (Adriaensen et al. 2019, Houghton et al. 2015; Lane et al. 2019; Revell et al. 2019; Stanton and Harvey 2017).

Systems ergonomics methods provide a highly useful means to analyse complex systems; systems which exhibit a number of key aspects (see, for example, Cilliers, 1998). Systems ergonomics methods can model many aspects of complexity such as interdependence, non-linearity, and the hierarchical nature of system and sub-system organisation. A criticism of these approaches, however, is that they provide only static descriptions of what are dynamic problems and systems. This means that analyses are often unable to consider key concepts of complexity such as emergence and dynamic feedback loops. Moreover, it is not possible, at least with current systems ergonomics methods, to dynamically simulate the behaviour of systems over time. To fully deal with the complexity inherent in the systems in which we work, the ergonomics discipline must begin to embrace methods that are able to model the dynamic nature of complex systems.

Computational modelling and simulation approaches such as System Dynamics (Sterman 2000) and Agent-Based Modelling (Bonabeau 2002) go beyond current systems ergonomics methods by providing the capacity to dynamically simulate the behaviour of complex sociotechnical systems. These methods work by identifying key variables that influence behaviour and using equations to define the attributes of variables, and relationships between them, to support computer-based simulation of system behaviour over time. As such they could provide ergonomists with the capacity to simulate both complex ergonomics problems and the likely impact of ergonomics solutions. Whilst these methods have a strong history of application in the social sciences, they are only recently being employed by Ergonomists. As such, there is an emerging body of work in which computational modelling approaches are being used to tackle ergonomics problems such as workplace injury (e.g. Perez et al. 2014), transport safety (e.g. Thompson et al. 2017), healthcare (e.g. Ibrahim Shire et al. 2018), emergency response (Baber et al. 2013) and sports injury (e.g. Hulme et al. 2019).

In a previous special issue of Ergonomics on the topic of Sociotechnical Systems and Safety, Hettinger et al. (2015) called for the adoption of systems ergonomics modelling and simulation approaches to improve our capability to design and evaluate systems in relation to safety and accident prevention. This call has since been echoed by others working with systems ergonomics methods (Salmon and Read 2019; Ibrahim Shire, Jun, and Robinson 2019). A central argument of these calls relates to the capacity of simulation techniques to evaluate proposed system designs, or proposed interventions, by identifying unintended consequences of design decisions.

This special issue on computational modelling in ergonomics is a response to these calls. There are three aims. Firstly, to provide a platform for communicating contemporary ergonomics research involving the use of computational modelling approaches. Second, to showcase the capacity of these approaches to dynamically model complex ergonomics problems and to inform the development of solutions. Third and finally, to inspire the ergonomics community to pursue further applications involving these approaches.

Overview of computational modelling methods

System dynamics modelling

Understanding societal dynamics is an essential prerequisite for developing effective policy. Therefore, methodologies to support societal level inquiry are critical. System dynamics provides is one such methodology, the outputs of which can be used as a policy tool within organisations or at the wider societal level. Traditional static quantitative approaches are not well suited for this purpose as they cannot deal with the multiple interacting relationships inherent in influencing outcomes, nor the dynamic nature of larger sociotechnical systems.

System dynamics is designed to deal with growth and decay in systems with interdependent components. Casual loop diagrams are generally developed as a starting point for the analysis. These identify the positive and negative feedback loops between key variables that influence behaviour over time (Sterman 2000). Positive feedback loops are reinforcing while negative feedback loops create balance. Key feedback loops can then be selected for more formal quantitative modelling in the form of ‘stock and flow’ diagrams. In these models, stocks represent accumulations of an entity (e.g. wealth) and the flow is the rate of change of the stock. Equations sit behind the diagrams to enable simulation of system behaviour over time. Empirical data can be used to input to equations, or as a benchmark against which to evaluate the model’s behaviour. Once the model has been validated for its intended use, it can be used to simulate the impact of real-world interventions, making it a useful policy decision support tool (Sterman 2000; McClure et al. 2015).

While it offers compelling benefit, as with all methods, considerable care needs to be taken in selecting this method only for the situations that require it. For example, system dynamics tends to operate at a high level of analysis, and aggregates entities into homogenous buckets. While this is an attribute when used in policy decision support, it become disadvantageous if the focus of the research requires delineation of individual differences. Taken as a whole however, and when used correctly, system dynamics is a compelling quantitative approach to support collaborative learning (Sterman 2000).

Agent-based modelling

An approach able to address some of the concerns associated with the lack of heterogeneity represented by populations in system dynamics models described above, is agent-based modelling (ABM). ABM is a true bottom-up approach meaning that while macro-level system outcomes can be analysed, ABMs are built at the unit of individual entities who interact and grow phenomena of interest. That is; populations are not modelled, but the effect of thousands of interacting people are; economic growth is not modelled, but the effect of competitive trading behaviour between businesses is; hospital infections rates are not modelled but the effect of interaction between health-care workers and community members under various hand-hygiene and hospital spatial configurations is. In this sense, ABMs are most useful for ergonomics (or other) researchers interested to explore not just of what occurs, but how and why.

ABMs are hugely diverse and there is some consternation over where their definition begins and ends (e.g., see individual-based models, discrete event simulation, micro-simulation). Most likely, they can be identified as models containing stochastic processes and where decision-making and consequent behaviour of modelled agents lies within the entity, itself and is not imposed by the model builder. The builder can create the rules of engagement, but the population must otherwise be ‘set free’ to act as they please. In this sense, ABMs have the capacity to incorporate a great deal of complexity in their design and outputs. However, the most elegant and enduring examples are arguably those that have combined simple behaviour and interaction rules between agents to produce complex and sometimes unexpected macro-behaviour that explains a phenomenon (Epstein 2013; Schelling 1971). Indeed, it is not at all unusual for ABMs to create behaviour that is totally at odds with what the model builder originally imagined might occur. This ability to uncover emergent, and even counterintuitive results is a key benefit of applying ABM (Bonabeau 2002). The challenge then is to determine whether the observed patterns are reasonable or not.

Discrete event simulation

As noted above, approaches like discrete event simulation (DES) are similar to ABM, and provide representations of some bounded aspect of the world. DES modelling is more focussed on details of process, and has been characterised as a process modelling approach. Like the other methods, DES modelling necessarily begins with the specification of a conceptual model. In specifying the model, choices are made about the aspect of the world that are modelled and questions are framed that further refine the focus of the model’s components. The model is then constructed using three main entities: system elements, resources available, and queues. These three entities are arranged in a structure that describes movement through a process over time (i.e. pathways). Data inputs parameterise the model. Model testing then occurs to validate the result as a useful representation of the world under study with respect to the question that has been framed. Then the model is ready for the simulation experiments, which mostly work to identify inefficiencies in the process and ways these can be overcome. DES tends to be used mostly to model service delivery processes, supply lines and manufacturing conveyors, where competition for resources, and timing of supply is critical.

Other computational approaches

Multiple other computational approaches able to deal at once with the availability of ‘big data’ and the complexity caused by dynamic interactions among people and entities are continually emerging. Some of these are analytical techniques needing to be ‘fed’ large swathes of existing data, and others (like ABMs) have the capacity to create their own. Many more incorporate cross-overs between methods and models that are at once both branching out and converging due to vast activity in areas related to artificial intelligence such as neural networks, reinforcement learning, genetic algorithms, graph analysis (e.g. social network analysis), and multiple other approaches. While commercial interests and/or computer science problems rather than ergonomics is the usual focus of these methods’ development, ergonomics researchers can benefit greatly from their introduction by considering how they might apply to our own questions. As computational methods mature and become more accessible, understood, and accepted in the ergonomics community, the more likely we are to realise their benefits – but also their limitations.

Applications of computational modelling in ergonomics

The contributions to this special issue demonstrate how various computational modelling approaches can be used to address a wider range of complex societal and work-related ergonomics issues across various domains from manufacturing and supply chains, healthcare, transport, and workplaces more generally. Davis et al. (Ergonomists as designers) open the special issue by presenting three applications of agent-based modelling and simulation. Their applications span the areas of engineering teamwork in the manufacturing section, evacuation behaviour of crowds and engineering supply chain processes. Practical insights from undertaking the three case studies are provided throughout, with a focus on the involvement of interdisciplinary teams and the participation of stakeholders within the modelling process. For example, in relation to crowd behaviour, the modelling identified emergent effects including doors creating bottlenecks for crowd movements, social contagion decreasing evaluation time, a lack of impact of falls on overall evacuation time and travelling in groups reducing evacuation time. Davis et al. conclude that computational modelling approaches offer ergonomics researchers and practitioners the opportunity to test ideas, interventions and design solutions at minimal cost in a way that is not feasible using real world methods.

In their application of system dynamics in healthcare, Farid and colleagues (Using system dynamics modelling to show the effect of nurse workload) demonstrate how system dynamics modelling can be used to simulate the effects of nursing workload on outcomes such as burnout, absenteeism and medical errors. This is a particularly relevant topic area as we go to publication, with healthcare workers internationally facing unprecedented workloads responding to the covid-19 pandemic. The modelling showed that long shifts and working weeks doubled fatigue levels, increased burnout and absenteeism and increased medical errors with patient safety impacts by up to 150%. This work is significant in showing how dynamic models can be used to demonstrate to management and policy makers the longer-term effects of their decisions on key outcomes of employee wellbeing and quality of patient care.

In a second application of system dynamics, Salmon et al. (Computational modelling and systems ergonomics) apply the approach in the road safety context to explore the potential impacts of different interventions designed to prevent drink driving-related crashes and fatalities. Salmon et al. developed a system dynamics model that simulates the behaviour of a drink driving-related trauma system. The model is subsequently used to explore what combination of road safety interventions would have the greatest impact in terms of reducing drink driving related crashes. Two types of policy intervention were examined: a “standard road safety” approach including conventional road safety interventions such as randomised breath testing and road user education campaigns, and a “public health policy” approach aiming to reduce the population prevalence of alcohol misuse in the community through public health measures such as taxation of alcohol, restriction of sales, responsible sales and service, and increasing diagnoses and recovery programmes for alcohol addiction. Based on simulations of the model over a thirty-year period, Salmon et al. report that the target of a 50% reduction in drink driving-related serious and fatal crashes over 30 years was only achieved by integrating the road safety and public health approaches.

Remaining with the road safety theme, Thompson et al. (The perils of perfect performance) investigate the potential consequences of introducing “perfect” autonomous vehicles into a system of imperfect humans who learn to manipulate Isaac Asimov’s first law of robotics, “A robot must not injure a human being”. The study explores how humans might adapt to the presence of autonomous vehicles in the environment and the difficulties this creates when distinctions between “manual” and autonomous cars are unclear or mistaken. Using an ABM, the paper discusses how the introduction of autonomous vehicles into a transport system where cyclists adapt to their behaviour by becoming less cautious can lead to a new source of risk where cyclists conflict with manually driven vehicles. The simulation also explored the effects on crash risk of different traffic mixes over the transition period from manual vehicles to autonomous vehicles. Overall, the model showed that although perfectly performing autonomous vehicles might reduce total crashes, they may create new sources of risk that offset some of the assumed safety benefits.

Returning to workplace settings, Antosz and colleagues (Employee shirking and overworking) present a computational model within an agent-based framework to explore employee behaviour, specifically how task distribution leads to underwork (“shirking”) and overwork. The model highlights the importance of the judgements of managers when allocating work. Managers must estimate task difficulty and thus how long the work will take to complete, as well as the competence of the employee receiving the work. They found that misjudgements of task difficulty gradually affects the accuracy of manager’s evaluation of employee competence levels, which can in turn lead to shirking and overwork. Given the increasing complexity of work, particularly for knowledge-workers, the insights from this model provide an interesting basis for how competing pressures in workplaces may be traded off under conditions of task uncertainty.

A deep neural network model was developed by Asadi et al. (A computer vision approach for classifying isometric grip force exertion levels) to address the concern of musculoskeletal injuries; a longstanding ergonomics problem prevalent across many industries. The paper describes the development and testing of a machine learning model to measure worker force exertion levels using videos of participant facial expressions and photoplethysmography (PPG; blood flow) values measured while they performed isometric grip exertions. The model development process involved extracting novel features from the facial recordings and PPG data, followed by training of the model using the participant data. Classification experiments and an activity task were conducted and the algorithm was found to have a good level of accuracy and to be robust to noise (i.e. participants speaking while performing the grip task). Asadi et al. suggest that the model provides an efficient means to provide accurate estimates of worker injury risk, without the need for physical observation of workers or disruption to work to enable measurements to be taken. The paper demonstrates how deep neural networks can be used to solve practical problems faced by ergonomics practitioners.

Next, Golightly and colleagues present an application of “multi-modelling” to unmanned aerial vehicle (UAV) control. They describe multi-modelling – a paradigm that enables integration of multiple computational modelling formats – and then describe how ergonomics theory, principles and methods were used within a multi-modelling approach to explore the ergonomics considerations relating to human supervision of multiple automated UAVs. The authors show how human performance characteristics were incorporated into a baseline model which focussed on the UAV performance only. This involved modelling the task activity of the human operator, operator occupancy levels and task switching. Further adaptations of the model were demonstrated around dynamic aspects of workload (performance penalties applied during periods of underload and overload) and windspeed (affecting UAV travel time and time spent at waypoints). The authors also discuss further use cases where ergonomics could contribute within a multi-modelling paradigm including in smart manufacturing, eco-driving in rail and crowd behaviours during evacuation.

Returning to the healthcare domain, Ibrahim Shire et al. focus on the usability and utility of the system dynamics method from the perspective of the stakeholders engaged in the modelling process. Specifically, they investigate the applicability of a participatory system dynamics approach based on the opinions of participants (hospital pharmacy dispensers and managers) in relation to usability and utility. They found that engaging stakeholders in a participatory system dynamics approach enhanced team learning through helping participants to develop a shared mental model, and by aiding decision-making and identifying trade-offs. The evaluation reinforced the benefits of stakeholder participation in model development (rather than testing or review), giving them ownership of model and thus higher confidence in the validity and utility of its outputs.

In the final paper of the issue, Holman and colleagues challenge us to consider the future of ergonomics practice and how our methods will need to change and adapt to address upcoming challenges of the Fourth Industrial Revolution characterized by technologies such as robotics, artificial intelligence, biotechnology and quantum computing. According to Holman et al. the systems of the Fourth Industrial Revolution will be more densely connected, more complex and be on a larger scale than that seen in the past. It is argued that these challenges exceed the capabilities of existing ergonomics methods. Their Radical Systems Thinking in Ergonomics Manifesto calls for the adoption of a “radical systems thinking” approach to ergonomics practice, the exploration of how computational modelling methods can supplement or augment current systems ergonomics methods, and the development of new methods capable to address problems in these future systems.

Barriers to the uptake of computational methods in ergonomics

As evidenced by the contributions to this special issue, computational modelling approaches can go beyond the capabilities of existing systems ergonomics methods to dynamically simulate behaviour. Importantly, many of the contributions confirm previous assertions that they can be used in an integrated manner with systems ergonomics methods (Salmon and Read 2019). It is our view that computational modelling approaches should therefore be embraced by the ergonomics community. This, however, is perhaps not straightforward to implement. While it is clear from the contributions that computational modelling approaches are beginning to be used to explore ergonomics issues, there appear to be some strong barriers to their more widespread adoption.

Several of the papers in this special issue specifically discuss barriers to the uptake of computational methods in ergonomics. For example, Davis et al. (this issue) note that there is a lack of awareness within the ergonomics community of computational methods, and a lack of the skills required to apply the techniques. This type of modelling is not a part of standard qualifications and training for ergonomists. Indeed, Salmon et al. point out that these approaches are more complex than systems ergonomics methods such as CWA (Vicente, 1999), STAMP (Leveson 2004) and FRAM (Hollnagel 2012). In particular, they require skill in identifying key variables and in designing the equations that enable dynamic modelling. Therefore, researchers and practitioners wanting to use these approaches need to invest time into developing the knowledge and skills to be able to create good models and use them appropriately; or build a collaboration with an expert in the technique of interest. As pointed out by Davis et al (this issue), this is important to protect against poor quality modelling being undertaken by those without appropriate training or without collaboration with specialist modellers. Such work could hinder the uptake of these approaches in ergonomics more widely, particularly with industry funders.

Another barrier is that the development and validation of computational models can be more time consuming than the use of traditional ergonomics methods, particularly if adopting participatory approaches with stakeholders (Ibrahim Shire et al., this issue; Salmon et al, this issue). Time requirements need to be considered in planning for these approaches to ensure that high quality engagement is possible, and that valid models are produced. However, this is not likely to involve a more significant investment of time than the use of existing static systems ergonomics approaches.

A final barrier relates more specifically to the reliability and validity of the outputs of computational approaches (Salmon et al., this issue). In considering this aspect, it is important to have clarity on the purposes of the modelling application. Potential aims may be the development of theory for further testing, the development of shared mental models of a system, the development of a means to refine design decisions, or a more formal evaluation of interventions to identify both positive effects and negative, unintended consequences. As these approaches become more commonly applied within ergonomics it is likely that discussions around reliability and validity requirements for different modelling purposes will become more critical. On the whole, model inputs can derive from established theory or empirical data, and outputs can be judged by their alignment with real world phenomena on key variables. Subject matter expert validation, commonly used to validate the outputs of systems ergonomics methods, can be another tool to ensure the quality of the model produced. Again, these issues are often raised in relation to ergonomics methods more generally (Stanton 2016), and while important to consider, should not prevent their uptake.

The future of computational approaches in ergonomics

It is our view that computational modelling presents an exciting new area for ergonomics. Although barriers exist, they can be overcome by improving awareness of modelling, training of researchers and practitioners in modelling techniques, formal reliability and validity testing, and through the use of interdisciplinary teams involving ergonomists working closely with more experienced modellers.

However, as with other ergonomics methods and approaches, computational modelling approaches should be considered another tool in the ergonomics toolkit, to complement existing methods and approaches, rather than replace them (Davis et al. this issue). Further exploration and guidance around when and how these approaches are best used would be a useful avenue for future research.

Importantly, ergonomists have a lot to offer to applications of computational modelling. Our understanding of human, team and system performance, as well as our theories, principles and existing methods can add an important evidence-base to model development processes. Further, work to integrate existing ergonomics methods with computational approaches is needed to trade off the strengths and weaknesses across methods. The integration of computational modelling approaches with other ergonomics methods has been proposed to offer a useful approach for extending the explanatory power of ergonomics methods whilst at the same time strengthening the validity of computational models (Salmon and Read 2019). Integrated approaches were demonstrated in the multi-modelling approach described by Golightly and colleagues (this issue) and further discussed by Holman and colleagues (this issue). Indeed, the Radical Systems Thinking Manifesto provides a set of capability requirements for future ergonomics methods whereby such methods must be capable of: (1) supporting allocation of function by modelling interactions between humans and autonomous agents and the effects of these interactions on the whole system; (2) modelling emergence to forecast and address latent risks; and 3) providing multi-level analysis to deal with large scale sociotechnical systems (Holman et al., this issue). Future methods development and methods integration, drawing on computational modelling to address the dynamic phenomena within sociotechnical systems is needed for ergonomists to continue to provide effective solutions to the complex problems we are often called upon to address.

Conclusions

As well as demonstrating the utility of computational modelling methods, the contributions to this special issue have provided a number of key findings. Some of the practical findings from the contributions included that:

• Variables such as falls, social contagion and travelling in groups can have un-intuitive effects on crowd evacuation times.

• Requiring nurses to work long shifts and long working weeks increases fatigue levels, and subsequent leads to increased burnout, absenteeism and increased medical errors with patient safety impacts by up to 150%.

• Integrated road safety and public health policy interventions are likely to be more effective than standard road safety approaches in reducing drink-driving related crashes and trauma.

• Introducing highly reliable autonomous vehicles into a transport system where humans adjust their behaviour to the lower risk posed to them could create new sources of risk that offset some of the assumed safety benefits.

• When managers underestimate the difficulty of a task and thus give shorter deadlines, they will more quickly gain an accurate understanding of employee competence than if they overestimate task difficulty.

• Machine learning models can be used to assess ergonomics issues such as high or repetitive force exertions in the workplace.

• Multi-modelling approaches incorporating human performance characteristics are useful to explore the effects of operator strategies and environmental factors on UAV system performance.

• Stakeholder participation and engagement in computational model development helps to ensure model validity, as well as promoting stakeholder learning and the development of shared mental models.

• Ergonomics requires new methods or combinations of methods to meet the challenges posed by the Fourth Industrial Revolution.

Ergonomics has a noble, yet challenging, mission. Paraphrasing the International Ergonomics Association’s definition – our role is to understand the composition and behaviour of systems in order to optimise human wellbeing and overall system performance. Our discipline takes a pragmatic and collaborative approach and we are constantly searching for new and improved ways to improve systems. Computational modelling approaches can provide new insights on the dynamic factors at play. They enable us to more fully examine and investigate the complex and dynamic nature of system phenomena. Equally, the ergonomics discipline has much to offer in terms of theory, principles, data and existing methods that can inform and improve the way in which computational models are designed. This special issue has showcased a series of applications where computational modelling techniques have been successfully applied to understand and respond to complex ergonomics problems. We hope to see more examples where these areas of expertise are brought together through cross-training and interdisciplinary collaboration. Such an integration can only help to improve the design of environments in which people live and work.

Acknowledgements

Gemma Read’s contribution to this work was funded through her Australian Research Council (ARC)Discovery Early Career Research Award (DE180101449). Paul Salmon’s contribution was funded by his ARC Future Fellowship (FT140100681) and Jason Thompson’s contribution was funded through his ARC Discovery Early Career Research Award (DE180101411).

Disclosure statement

No potential conflict of interest was reported by the author(s).