1. Introduction
Socio-technical systems (STS) exhibit technical and social complexity and hence are more difficult to specify than software or hardware systems alone [
1,
2]. The technical and social parts of STS have complex interactions and thus the identification of suitable specifications is far from easy [
3,
4]. Capturing requirements constitutes a central activity in STS design and implementation. Failing to thoroughly capture and validate requirements is a key reason for system failure [
5,
6]. During requirements capture, designers need to decide on the most suitable level of automation, also referred to as functional allocation (FA). During this activity, technology should be viewed as a tool to assist humans to meet their goals, rather than implemented because of assumed efficiency or cost-savings [
7]. Functional allocation and other human factors issues, that refer to technological, environmental and organisational aspects that influence human performance, are rarely modelled and linked to requirements in STS design [
8]. The need to address human factors in STS has been highlighted in domains such as healthcare [
9], transportation [
10,
11], military [
12], city design [
13], policy design [
14] and business [
15] with [
16] emphasizing that designers should consider human psychology, throughout the STS design. Methods that address STS requirements focus on work processes [
17] and apply formal modelling such as REASSURE [
18], which may utilise input from experts, although they do not explicitly address human factors.
Practitioners in STS design [
16,
19,
20], have requested methods that are not just descriptive but explanatory or predictive in nature and with the ability to test the integrated human activity and task-support through computer-based models and simulations, such as system dynamics [
21,
22], agent-based modelling or digital twins in virtual settings [
23]. These techniques however have not been used effectively for designing STS with human factors in mind. New simulation approaches are required to link the top-level aspects of systems with low level specifications that support human factors concerns [
21]. Evidence from the application of two popular deign methods used by human factors experts [
24], the cognitive work analysis and its successor, the cognitive work analysis design toolkit [
7], highlight that software tools, simulations and computer-based modelling are needed for evaluating the effects of different designs. In this paper we address this limitation through the application of VR simulation while also introducing a new requirement to bridge the gap between the human and technical facets of STS. We define these as Task Support Requirements (TSRs) to explicitly describe how technology can support human activity (tasks) and performance while addressing human cognitive limitations. TSRs aim to also provide a ‘lingua franca’ for software engineers and HF experts to discuss requirements that relate to functional allocation and other HF issues.
The proposed method utilises virtual prototyping based on [
25] and it is related to simulation-based requirements validation methods [
26,
27,
28], that utilise Bayesian networks and evolutionary computing to validate non-functional requirements (NFR) and optimise requirements specifications in complex STS. Alternative methods such as physical prototyping, could be used to test TSRs, but are expensive to implement whereas simulated environments can reduce validation costs, especially for complex systems [
29,
30]
The contribution of this paper is a new STS method that incorporate TSRs as a representation to bridge the gap between what people do (Tasks), and what the computer will provide (Functional requirements) and the shared user interface. The method combine existing requirements engineering notations (i*, Goal trees-GORE, and Design Rationale) through a framework for considering design alternatives that are influenced by human limitations. The problem addressed is the lack of methods that explicitly consider human factors when specifying requirements to support human limitations. Unlike other STS approaches such as, [
19,
31,
32,
33,
34,
35,
36] that are based on conceptual models or address functional allocation in a limited manner, the proposed method aims to optimise human activity while validating solutions experimentally by virtual prototyping.
The proposed method is evaluated using case study methodology in two phases. The first phase provides a detailed application of the method during the early stages of designing a smart in-vehicle information system (IVIS). The second phase provides an empirical evaluation with expert and novice designers in specification of a road planning application to enhance pedestrian safety. The research questions addressed are:
The paper is organised as follows. First, we review the literature on STS design, requirements analysis approaches and human factors issues: situation awareness (SA), workload and functional allocation. Next we define TSRs and the proposed methodology that utilises TSRs. A detailed case study is presented showing an application of the method during the design and validation of an in-vehicle information system. Next, the empirical evaluation of the method is presented using a different case study (risks from contagious diseases during pedestrians commuting). The paper concludes with lessons learned, discussion of the implications of this method and the findings.
2. Related Work
Methods for STS design attempt to elicit user needs either through understanding the problem or designing an optimum solution given the properties of constituent system parts. ETHICS [
37,
38,
39] and QUICKethics [
40] claim to give the same attention to the needs of the people involved as to the demands of the technology; however they have been criticised for being slow and costly, with the involvement of unskilled users in the design process [
41], and lacking tool support [
42]. Hickey et al. [
43] tried to integrate ETHICS with agile approaches such as Extreme Programming, Dynamic Systems Development Method, and Scrum [
44] which incorporate user involvement to address user needs; however, agile approaches are mostly concerned with end-user requirements with no reference to human factors that inherently affect user performance. Soft Systems Methodology [
45,
46] takes into account stakeholders’ differing viewpoints to solve a defined problem, but also ignores human factors. Cognitive Work Analysis [
36,
47] aims to predict what a STS could do, and refers to actors’ cognitive skills but not their cognitive limitations. Cognitive systems engineering [
34,
35] deals with the analysis of organisational issues based on human factors; however, it lacks the technical systems design dimension. Human-centred design [
48] is based on understandings users’ needs and requirements and explicitly refers to social and cultural factors, including working practices and organisational structure, by applying human factors/ergonomics, and usability knowledge and techniques. The main criticism of this method is that the analysis tends to view human activities as a static sequential process [
49]. The System-Scenarios-Tool is a user-centred methodology for designing or re-designing work systems that uses human and machine properties. Its main limitation is that is largely a conceptual method without tool support for modelling and simulation [
4]. Other systems engineering methods for STS design include Adaptive socio-technical systems [
31], which use a requirements-driven approach to self-reconfigurable designs using Tropos goal modelling; and the Functional Resonance Analysis Method(FRAM) [
19] which is based on resilience engineering and analyses possible emergent behaviour in complex systems.
Overall, STS design methods use evaluation tools based on static or simplified conceptual models or mock-ups that do not explicitly take into consideration how human factors should be addressed during the functional specification of interactive software. However, the majority of human factors analyses investigate the human factors alone and not how they can be used to specify solutions to support people working practices [
50,
51,
52]. These shortcomings highlight the need to improve STS to address the complexity of human-system interaction [
26] and to optimise level of automation, (functional allocation) [
53,
54,
55], that could be in any of the eight different levels defined in [
56]. When allocating tasks between the human operators and the automated system, inefficient automation design often arises from a lack of consideration of the role and limitations of human operators and of their interaction with the automated system [
57]. Early FA methods such as the Fitts heuristics [
7] aid allocation of functions between human operators and machines, by defining tasks that machines tend to perform “better” than humans and those that humans perform “better” than machines. Fitts suggested that machines perform better routine tasks that require high speed, force and computational power, while humans undertake tasks that require judgment and creativity. They also acknowledge the limitations of humans in correctly employing these capabilities when overloaded with excessive task demands, or maintaining alertness due to fatigue. Fitts’ MABA (machines are best as) list, despite its age, has persisted throughout the history of functional allocation [
58].
One strategy is to increase automation and design out human error; but this comes with its own penalty of impoverished situation awareness (SA), the ability of a human agent knowing what is happening around him/her. This in turn leads to subsequent errors from leaving the human agent out of the loop [
52]. The use of software in safety-critical areas such as intelligent transport systems has increased significantly, so software failures can impair system safety [
59]. This highlights the need for better allocation of functionality between human and technology. Results from the analysis of accident causality indicate that more rigour is needed in analysing HF requirements in safety-related systems [
60]. Inadequate or misunderstood requirements [
61] relating to HF is a major cause of accidents [
60]. Methods for partitioning functions between automation, human-only operation and cooperative human-computer functions have been proposed in Human Computer Interaction [
62] and need to be addressed explicitly and strategically at an early stage of STS design to maximise chances of success [
8].
We argue that requirements analysis should incorporate FA to specify software requirements to support human tasks, capabilities, and skills. Previous work such as task descriptions [
63] define what user and system must do together [
64] using problem space analysis to identify requirements [
65]. Work on the integration of goal-oriented with problem-oriented requirements engineering address a wider scope of the to-be system [
66], however, they fail to address the human factors part that need to be supported to minimise STS failure.
Human factors concerns has been partially addressed in i* modelling [
32] through skills and human agent capabilities and goals-skills matching [
33]. However, i* does not address mapping human activities and capabilities to system requirements that support human action and cognition (SA, workload, etc.). Past NFR frameworks [
67] with Softgoal Interdependency Graphs [
68] using the i* notation [
32], have addressed issues such as reliability/performance, however criteria for their satisfaction are judged without reference to Human factors.
In [
69], authors use Quantified Softgoal Interdependency Graphs (QSIGs) to assess the degree of softgoal satisfaction. However, the assessment is based on subjective estimates of the degree of interdependencies among soft goals. Virtual prototypes [
50,
70,
71] have provided designers with multiple viewpoints of the system’s functionality that assists in requirements validation, e.g., the Immersive Scenario-based Requirements Engineering method [
25]. In the automotive industry, VR (virtual reality) is used to test the safety of a vehicle while minimising design costs [
72]. The advantages of VR and simulation however have not been fully leveraged for STS design due to the complexities of such systems.
5. Detailed Application of the Proposed Method
A case-study illustrates an application of the STS design method in the context of smart in-vehicle information systems. The aim is the design of a STS to support drivers’ situation awareness (problem). The process starts with the specification of drivers’ information needs in terms of goals, identifies the optimum distribution of tasks between the driver and potential software technology (functional allocation) to address these needs, refines the selected FA option into TSRs, and validates the TSRs through VR simulation.
Step 1. Problem specification: Driver safety & support systems
Design of In-Vehicle Information Systems (IVIS) and Advanced Driver Assistance Systems (ADAS) to assist drivers with the complex demands associated with the driving task [
84] have explored technologies such as lane departure warning, lane departure prevention, active lane keeping, front crash prevention, blind spot monitoring, rear-cross traffic alert and driver monitoring systems [
85]. Automotive design guidelines describe desirable practices that are not mandatory and hence are less strict than standards [
86].
In traffic safety, situation awareness (SA) and workload constitute critical safety factors as associated Non-Functional requirements. Situation awareness enables the driver to anticipate events under the perceived driving and environmental conditions [
87], and is defined as the process of perceiving information (level 1) from the environment, comprehending its meaning (level 2) and projecting it into the future (level 3). This is linked to the 3-level model of driving (operational/tactical/strategic) [
88], referring to actions for stable control, manoeuvring and route planning. Work by [
89,
90] stress that operational driving tasks such as steering and braking responses primarily require level 1 situation awareness support, although level 2 situation awareness may also be involved [
89]. For IVIS to improve drivers’ situation awareness it is essential to enhance their ability to perceive and interpret traffic and environmental information (situation awareness levels 1 and 2) to support the tactical and operational tasks of driving.
Notifications can assist drivers’ tasks and changes in their environment [
91] however, design of effective notifications is challenging [
92] since notifications can also be distractors. In the same vein, workload is linked to situation awareness and refers to the limited cognitive resources of humans and how this can affect human reliability. So, if a hazardous situation emerges when the driver is overloaded, the risk of committing an error is increased.
Step 2. Goal modelling and high-level TSR specification
Goals in this model refer to the 3-level model of driving tasks: strategic, tactical and operational [
88]. The first is associated with strategic driver decisions and tasks that relate to the selection of the best route to arrive at the destination. The criterion here could be travel time, scenery, etc. At the tactical level, goals are associated with actions of the driver that relate to the desired manoeuvres to achieve short-term objectives such as overtaking. At the operational level are goals relating to manoeuvres to control the basic operations of driving such as acceleration, deceleration (speed control) and lateral deviations (direction control).
Figure 2 depicts the goal hierarchy graph.
During this step goals are decomposed to a level where assumptions about automation become apparent, such as “control vehicle” in contrast to a non-automated solution such as walking. At this stage TSR become apparent for achieving lower-level goals. For “control direction” and “speed”, task support is delivered by standardised controls of steering wheels, brake and accelerator pedals, although further decomposition and definition of TSR is possible; for instance, in cruise control for “control speed”. In the case of cruise control the user interface implications are refined into status displays and controls to set/disengage cruise control mode.
Step 3. STS modelling using i*.
The i* model of
Figure 3 provides the link between goals in
Figure 2 and softgoals (NFR) that need to be satisfied by functional requirements for the system to be successful. The focus in this step is “Monitor environment” and “Respond to hazards” sub-goals (
Figure 2). Monitor environment, depend on the softgoals “Maintain safety” and “Maintain situation awareness” in
Figure 3. The “Respond to hazards” goal also depends on the NFR “Maintain situation awareness” and is decomposed into the tactical, operational tasks of, halt, avoid and warn (pedestrian risks and other road users). These tasks can be supported by technology/functionality (depending on desired level of automation explored in subsequent steps) which in return will satisfying the associated NFR/softgoals “Maintain situation awareness” and “Maintain Safety”. For instance, in the case of system warning, the specification could be refined to provide only critical information of traffic conditions to the driver with an audio warning of imminent threats that are not visible.
Step 4. Functional Allocation analysis for “automated warning” option and HCI modality analysis
In this case study we illustrate the rational for selecting the automated driver warning option through functional allocation (FA) analysis, selection of the most appropriate HCI modality, and the refinement of this option into low-level tasks (Figure 6) and TSRs (
Table 1).
The FA approach we adopted is based on the combination of the Fitts model and the automation taxonomy framework of [
81]. Parasuraman’s framework is mapped on to different stages of human information processing: 1) information acquisition, 2) information integration (comprehension) 3) decision making, and 4) response. Automation can operate at varying levels in all stages of information processing.
Figure 4 illustrates a high-level functional allocation analysis for the “respond to hazards” goal as a design rationale diagram where the goal corresponds to the question asked. The figure illustrates two functional allocation options: “manual” and “automated hazard recognition and response”, with the former being rejected as it provides no support for SA. The Automated option is decomposed into three more detailed options: 1) complete automation for recognition and hazard avoidance which would require considerable AI processing and is currently being developed in driverless vehicle technology; 2) automated halting which relies on AI; and 3) automated warning of the driver with speech/audio or visual display which does not depend on AI technology.
The AI option of automated halting and avoidance would be the most expensive; however, all options depend on some automated processing to detect dynamic hazards, i.e., other vehicles and pedestrians. If the automated halt/avoid technology works reliably it would be the safest option, but reliability and security doubts [
93] reduce this advantage [
94]. Warning the driver of hazards contributes to safety and situation awareness with a lower cost and better reliability. This option is selected as it represented the best criteria (NFR) trade off. The warning option is then decomposed further to investigate different HCI modalities, audio/speech/haptic warnings or visually through a visual situation display as depicted in
Figure 5. The speech option may encounter reliability difficulties in giving precise instructions and location of the hazard within the microsecond time scale. Furthermore, the driver may not have the necessary mental map of the situation to execute immediate response, so situation awareness is not supported. The same applies for audio messages such as beeping from different orientations within the vehicle. The option “Visual warning” does support situation awareness and should encourage the driver to maintain a mental map and awareness of the road situation and potential hazards. Therefore, a “
Visual situation display” option that provides support for situation awareness is chosen as the best solution.
Step 5. Decomposition of the automated warning via “Visual situation display” task into sub-tasks that need to be performed to maintain adequate situation awareness.
The selected visual warning option is refined into low-level tasks showing key activities the driver needs to perform to maintain sufficient level of situation awareness (see
Figure 6). This analysis depends on domain knowledge, in-vehicle systems design literature and driver information needs [
95]. The functional allocation decisions (
Table 1) for each task need to specify which of these tasks should be supported by the visual situation display and what level of automation is appropriate for each task.
Table 1 illustrates summary of the trade off issues.
Step 6. Functional allocation analysis for the sub-tasks of the “Provide visual warnings” task and specification of functional requirements to support these tasks.
Functional allocation analysis of tasks in
Figure 6 is performed in tabular notation as shown in
Table 1. This is used as an alternative to QOC diagrams when the number of options and criteria combinations is large. The evaluation of each activity in terms of reliability and automation capability is assessed on the scale of high, medium, low (H/M/L). High indicates that technology is judged to provide superior results to human operation, hence full automation of the activity is possible with current technologies. Low indicates that people are better at performing this activity than the available technology and this task should be allocated to the human.
Tasks that are suitable for human-machine collaboration are specified in terms of task support requirements (TSRs) for interactive user interfaces, while the human-only tasks become manual operating procedures. The TSR specified in
Table 1 (rightmost column) refer to the visual situation display option that is based on automated warnings and visual HCI modality selected in previous steps. TSR are further analysed in the following design rationale step (Step 6a), where the design specification of candidate options becomes more apparent (Step 6b). In a similar manner the “maintain optimal workload” goal can be refined into its TSR and analysed for functional allocation options.
Steps (6a and 6b) describe the transition from the general method aimed at specification into the design phase where domain specific reasoning and trade offs are explored. This detail is given in
Appendix B which reports further design rationale analysis that produces two preferred options (radar/arrows) which are then subject to validation studies using virtual prototyping in the final stage (step 7,8 presented next).
Step 7. Implementation of virtual prototypes of the Radar and Arrows designs based on selected TSRs and specification of NFR metrics and VR scenarios.
Virtual prototyping is used to determine which of the two candidate designs will be optimal under a range of operational conditions. An experiment was conducted with participants using a VR driving simulator that incorporated the candidate designs of Arrows and Rarar (see Figure 9—
Appendix B). Designs are evaluated against situation awareness and workload NFRs.
The VR simulator is customised to create a replica of the environment and hazard scenarios that drivers are likely to experience, to simulate increased workload and stress their situation awareness. Three steps are followed during VR customisation: 1) development of the test traffic environment in terms of buildings, infrastructure and traffic flow; 2) model scenarios in terms of traffic flow and hazards; and 3) modelling the candidate designs through head-up display (HUD) technology. The HUD designs were specified from the requirements refinement process and the design rationale steps in
Appendix B, producing the virtual prototypes (Figure 9).
Virtual prototyping may require input from HF experts; however, domain analysis should provide scenarios, and the design rationale trade off criteria become measures in the experiment. During the experiment, driving behaviours were monitored and logged into the simulator’s database. The logged observations from the simulation were analysed to represent performance data (i.e., driver errors, potential accidents, perception of hazard-critical information) to select the design that best satisfies the NFR criteria. If the minimum level of the NFR criteria is not satisfied, then the virtual prototype needs to be redesigned and the process repeated until the NFR is satisfied.
To be confident that the design supported situation awareness, the situation awareness score was set at >= 60% indicating that the driver should be able to perceive 6 out of 10 separate critical information cues, which represents the minimum level of situation awareness required to maintain safe driving and is a quantitative estimate of the driver’s awareness of: vehicle(s) in blind spot, vehicle(s) ahead, behind and to the side of the host vehicle, pedestrians on road, obstacles, own speed limit, parked cars, congestion, position in the road-lane, distance from vehicle(s) ahead and behind. This threshold is based on Miller’s [
96] seven plus/minus two model and the useful field of view test, indicating the minimum information an individual can extract from a dynamic environment [
97] along with general driver visual information processing capacity [
98,
99,
100]. Workload NFR satisfiability was measured through an optimal rage of electroencephalography (EEG) scores in the range 45 to 70 out of 100, indicating the optimum level of workload under which the driver remains vigilant but not bored.
Step 8. Simulation-based validation of TSR based on selected NFR evaluation metrics.
Seventeen participants from the local population, with a valid driver’s licence and 20/20 vision or corrective glasses or lenses, took part in the experiment. Subjects selected had at least seven years’ driving experience and were under 55 years old. Prior to the experiment, they were screened for colour blindness or susceptibility to simulator sickness. They were introduced to the various simulator controls, made adjustments to the seat and were given a five-minute training session. In the before stage, subjects had to complete the Manchester Driving Style questionnaire [
51] to identify their driving style along with demographic information (average age was 37.1 years and the gender distribution was 55% female to 45% male).
During the experiment drivers were expected to drive along a pre-specified path in the virtual environment. The driving controls include a real steering wheel, brake and accelerator pedals, and a simulated automatic gearbox. Driver behaviour data was recorded including: lane deviations, headway (distance or duration between vehicles), speed, acceleration, EEG and deceleration. In total, 8,460 data-points were collected from each participant and each of the variables. The situation awareness assessment was conducted using the SAGAT technique (Situation Awareness Global Assessment Technique) [
87], by freezing the simulator at different points during the experiment and asking the participants to answer a number of questions that referred to the driving situation. Questionnaire responses from this process were analysed and assessed on a 0-100 score by comparing the actual situation with what the participants reported in their results.
Results from the experiment showed that both designs were significantly better than the control condition (no use of visual situation display). The required levels of the NFR criteria for both designs were satisfied, with drivers’ situation awareness level being on average 60% in all road sections. The two-way ANOVA repeated measure analysis that was carried out on the aggregated SAGAT score and the other dependent variables (speed, EEG and headway) for three data collection points that coincided with hazardous events, and three design conditions (radar, arrows, and control), identified a significant main effect for design on situation awareness (F(2,15)=10.90,
p<0.01). The radar design (mean 74.3) was superior to arrows (71.72) and the control condition (51.15). Both the arrows and radar designs were significantly better than the control (post-hoc tests,
p<.001) which verifies that the designs as specified by TSR satisfy the NFR “Maintain situation awareness” and “Respond to hazards” goal in i* model of
Figure 6. Thus, the design process ends.
6. Empirical Evaluation of the Proposed Method
A summative, qualitative evaluation of the method was conducted with 19 participants, 9 experts from the domains of information systems, intelligent transportation systems, and computer science, who had been working as professional systems analysts/consultants for more than seven years, 5 postgraduate students that recently completed a postgraduate course in e-business systems design and 5 novice participants with computer science background. The average age was 35 years, and the gender distribution was 63.2% male) The validation focused on the first 6 steps of the proposed method and aimed to evaluate the usefulness and correctness of TSR specifications that emerge while participants designed a hypothetical STS.
The criteria utilised to evaluate the method are based on [
7,
11,
22] and cover aspects pertaining to the method’s generalisability, learnability, effectiveness, usability, and support for human factors in the design.
A workshop was prepared in a domain that all participants were familiar with: minimise the risks from contagious diseases (e.g., COVID19 pandemic). Expert and novice subjects applied the method to design a new mobile application to assist travellers minimise their risk of conceiving a contagious disease (COVID19) while commuting in public spaces, by undertaking the activities in
Table 2 and answering the questions in
Appendix A. The evaluation was carried out in two phases, the first included a 3-hour session with students and novice subjects while the second phase involved 90 min individual sessions with experts. Subjects were initially trained on the methodology using the driver situation awareness example, and then asked to apply it to the COVID19 scenario. The goal was to specify the most appropriate functional allocation and specification of TSRs that could address one aspect of the COVID19 problem such as contact tracing/symptoms checking. Upon completing the exercise expert participants were asked to complete an online questionnaire (see
Appendix A) followed by interviews by the researchers. The interview was unstructured and started with open questions to elicit experts’ opinions about the method’s advantages and limitations. All interviews were recorded. Following the exercise, novice subjects were asked to complete an online questionnaire and asked to explain how they applied the method to come up with their design/TSR. Both experts and novices submit their completed questionnaires via Google forms and their designs via email. Participants’ designs/ were evaluated in terms of how well they contributed to the problem (COVID19).
The evaluation of results showed that experts perceived the method as easy to learn, structured, helpful in framing their thinking and efficient in addressing HF through the specification of TSRs.
Figure 7 show the percentage of subjects assessing each evaluation question with a score above the 3 in a 5-point Likert scale. These results indicates that the method can contribute positively to designing STS and address human factors issues effectively. The practical part of the evaluation was completed by experts, with >75% of participants scoring >65% in each assigned task as shown in
Table 2. The 65% threshold was selected to focus on participants with above the minimum acceptable level of performance (50%). Higher thresholds (e.g., >75% or more) were not used since they minimised the number of passing cases and constrained the knowledge to be drawn from these cases. The evaluation of each task was performed by examining the correctness of the produced outcomes with reference to the requirements of
Table 2. For instance, in the first task two example correct answers were “Find the best route to my destination with the minimum infection risk” and “Being aware of the infection risk at a given public place. Contrary to the experts, students and novice participants found the method more challenging possibly because of their limited knowledge in systems design. They primarily addressed functional allocation at the high level with limited attention to human factors. In contrast, the experts addressed the human factors in more detail and their designs had a strong link with the associated NFR criterion (contextual risk awareness).
Analysis of interviews with experts highlighted limitations and recommendations. Experts mentioned the need for tool support on functional allocation’s selection-criteria (what criteria should be used to decide for functional allocation) and possible software support to guide the exploration of the vast space of possible solutions. Experts recommended possible improvements: the use of taxonomy or advice on potential human factors limitations in different domains and tool support for design space exploration to assist in selecting the best UI options (modality, metaphor) from past similar systems using techniques such as analogical reasoning.
Overall, the evaluation of the method showed that it is useful in specifying requirements of STS to support human activity and addressing human limitations.
9. Discussion
The FA technique employed in the method stems from the HF literature and provides guidelines for the best allocation of tasks between humans and technology according to their strengths and weaknesses (Men are better/ Machines are better at- MABA/MABA) [
54]. Work by [
101], developed the task technology fit model to describe the optimum fit between managerial tasks and mobile IT capabilities under different environmental conditions, to improve overall task performance. Tasks are described in terms of routineness, structure, time criticality and interdependencies; while capabilities of mobile IT are seen in terms of functionality and user interface, and context in terms of distractions and obstacles. Their model, however, is specific to the mobile IT domain and requires further empirical research before application in other settings. Our TSR approach, is based on general cognitive theories, can be used in different disciplines with minor adjustments according to the available automation capabilities.
Some FA theories argue that the a priori allocation of functions as illustrated in Fitts’ List is an oversimplification [
34,
102,
103] claiming that capitalizing on some strengths of computers does not replace a human weakness. Instead it creates new human strengths and weaknesses that are often unanticipated [
73]. Dekker and Woods [
103] recommended that system developers abandon the traditional “‘who does what” approach of FA and move towards STS. Despite this criticism Fitts model remains popular for its generalisability and descriptive adequacy [
58]. Hence, in our method we utilise the Fitts List in accordance with Parasuraman’s model for the specification of an initial functional allocation.
Several STS design methods have been developed over the past 40 years, including ETHICS, QUICKethics [
37,
40], Soft Systems Methodology [
46], Cognitive Systems Engineering [
34,
35] and Human-Centred Design [
48]. However, most of these are rarely used [
20]; the main criticism being their limited capability in addressing prospective STS designs or providing evaluations, concentrating on problem analysis with existing systems rather than design solutions. An important issue in existing STS design methods is the different and sometimes conflicting value systems among stakeholders, such as improving job satisfaction and the work-life balance while at the same time achieving the organisation’s economic objectives. Empathic design [
104] and contextual design e.g., [
105] do consider the user’s environment as part of the development process, but their application has been limited. The STS method we proposed uses participatory techniques by involving users in the evaluation of the prospective system design. The experimental nature of the evaluation step encourages involvement of stakeholders (drivers in our example application).
Many STS methods have focussed on safety engineering involving diverse approaches such as Activity theory [
106]; cybernetics [
34]; Joint Cognitive System [
34]; Work Domain Analysis [
107]; Functional Resonance Analysis Method [
19,
108] and Framework for Slack Analysis [
109]. These are either techniques to address specific problems or are descriptive in nature and focus on showing how work is currently performed. Baxter et al. [
20] highlighted the inability of exiting STS methods to address prospective designs due to the difficulty of predicting the interaction between people, technology and context in a system world that does not exist (new world problem). Our method offers a solution to this problem through the introduction of TSRs that bridge the disciplines of HF and technology design and integrates existing modelling languages from software engineering and other disciplines.
TSRs extends previous approaches to STS design [
20] by providing a more detail-focused method that addresses the frontier between software design and higher-level heuristic design (human factors and goals) of STS. Mumford’s ETHICS [
40] containing general heuristics for analysing and shaping the components and human roles, within a framework of principles for human design of work practice, workplaces and organisations; however it does not address technology. A review [
20] of STS approaches, propose a research agenda for a system engineering approach to STS, oriented towards a high-level view of process and systems organisation. Similarly Design X, [
16] provides another high-level view on STS emphasizing system complexity, the role of people therein, emergent properties and the inherent complexity in STS. In contrast TSR provide a lower level design focus where components of human-computer activity could be considered within a higher-level framework provided by ETHICS and related approaches [
16]. Activity theory [
110] can operate at a similar level of granularity as TSR; however, it only provides a modelling framework of goals, objects, and activities without any view on functional allocation or definition of requirements.
The closer relatives of TSRs and our method are human factors oriented methods, such as Ecological Interface Design (EiD), [
36], CREAM [
111] and FRAM [
19] which focus on human safety engineering rather than functional allocation orientation of TSR. FRAM provides organizing principles and an activity modelling approach for analysing systems functions and their interfaces; however, as Hollnagel notes, it is a framework for problem diagnosis, rather than giving more detailed advice on FA, human factors guidelines and user interface design. Nevertheless, these methods could be used in conjunction with TSR. For example, the graphical representation as realistic metaphors of the system and software world for control interfaces, which is proposed in EiD [
36], could elaborate the user interface component of TSRs TSR draw upon human error frameworks [
51] and more general ergonomic advice [
112] to inform design and specification.
Requirements engineering methods e.g., [
113] have not addressed specification of software support for human decision-making and system operation, which form the focus of TSRs. Modelling of human agents and activities is presented in requirements engineering models such as i* [
114] which also support investigation of functional requirements (hard goals in i*) and NFRs (i* soft goals). However, i* modeling does not advise on design of user interface components or functional allocation. Modeling of requirements for adaptive systems [
31] provides detailed agent-activity-goal models using a formal extension of i*, combined with a ‘monitor- diagnose-reconcile-compensate’ framework for considering modification of user support requirements however task allocation and human factors advice is not supported. FRAM and TSR could be complementary with FRAM operating at the high system components level while TSR unpack system components in terms of software requirements, human operational activities and desired operational conditions. Investigation and validation activities using VR or simulation [
115] are time consuming, nevertheless, the low level of granularity employed VR simulation enable the quantitative evaluation of prospective systems, filling the gap in the existing STS design methods identified by [
20]. Guo et al. [
116] report a VR-based system to assist product design in its early stages. Through interacting with virtual prototypes in an immersive environment, the designer can gain a more explicit understanding of the product before its realisation. Their application of VR, however, is not linked to a systematic process of product design. Secondly, the development of high-fidelity VR prototypes can be prohibitively expensive. An alternative could be head-mounted VR displays, although users may be even more prone to motion sickness than in fixed-base [
72,
117].
Although it is intended to be generic, the method was tested in a specific context, and therefore generalizations about its effectiveness need further investigation. Furthermore, there may be significant differences in the level of complexity in STS, which may relate to challenges not identified in this study.