Connection Science, Vol. 14, No. 4, 2002, 283–295

Respective demands of task and function allocation on human–machine co-operation design: a psychological approach JEAN-MICHEL HOC* and SERGE DEBERNARDy *Institut de Recherche en Communications et Cyberne´tique de Nantes (UMR CNRS 6597), Centre National de la Recherche Scientifique, B.P. 92101, 44321 Nantes Cedex 3, France email: [email protected] tel: þ33 2 40 37 69 17 y

Laboratoire d’Automatique et de Me´canique, Industrielles et Humaines (UMR CNRS 8530), University of Valenciennes, France Le Mont Houy, 59313 Valenciennes Cedex 9, France email: [email protected] tel: þ33 3 27 51 13 73 Abstract. Co-operation between human operators and autonomous machines in dynamic (not fully controlled) situations implies a need for dynamic allocation of activities between the agents, in order to adapt the human–machine system to unexpected circumstances. Dynamic allocation is a way, for example, to avoid human workload peaks. Depending on whether tasks or functions are allocated, the demands made on human–machine co-operation design are different. Task and subtask allocation assume that both the human operator and the machine (or its designer) share the same decomposition of the overall task into subtasks. Function delegation is less demanding, provided that the human operator delegates functions to the machine explicitly, and within the context of a task representation transmitted by the human. This paper discusses these principles on the basis of experimental results taken from a series of studies on human–machine cooperation in air traffic control. Keywords:

human–machine co-operation, dynamic task allocation, dynamic function delegation, human–machine system design.

1. Introduction Dynamic situations are frequently encountered within highly complex and risky systems such as air traffic control, glass-cockpit aircraft piloting, nuclear power plants, and so on. They are not fully controlled by their human operators. The reason is twofold. Firstly, the environment is not fully predictable (e.g. a heading instruction to an aircraft can modify its speed because of the wind, thus generating an unexpected conflict with another aircraft). Secondly, not only human operators, but also autonomous machines are acting upon the same objects (e.g. in emergency situations, automatic devices are triggered without any front-line human operator intention). Connection Science ISSN 0954-0091 print/ISSN 1360-0494 online # 2002 Taylor & Francis Ltd http://www.tandf.co.uk/journals DOI: 10.1080/0954009021000068745

284

J.-M. Hoc and S. Debernard

As Hollnagel and Woods (1983) have pointed out, modern human–machine systems (HMSs) should be considered as joint cognitive systems. HMS design implies the definition of activity allocation between humans and the devices they are using. Fitts’s (1951) seminal work on principles for a priori allocation (The Human is best at . . . The Machine is best at . . . ) has been widely criticized because of its restricted adaptive power (e.g. Bainbridge 1987, Parasuraman 1997, Hoc 2000, and many others). For a long time, human engineering has expended some effort in defining dynamic allocation principles in real time (e.g. Rieger and Greenstein 1982, Millot and Mandiau 1995). Optimization parameters for such a dynamic allocation are diverse, e.g. human workload (maintained between two boundaries in order to avoid error-prone overload on the one hand and boredom underload on the other), accessibility to data (the agent who has easy access to the necessary data will do the job) and so on. In recent literature, the real time allocation principle is referred to in various ways, but using similar concepts—dynamic task (or function) allocation (Older et al. 1997, McCarthy et al. 2000) or adaptive automation (Kaber et al. 2001). Although these concepts place emphasis on the adaptation of the machine to the human, the studies also integrate the adaptation of the HMS to its environment, evaluated by looking at cognitive costs (for the human), performance (overall quality of results) and risk management quality. In this paper, we shall consider two extreme dynamic allocation modes—dynamic task allocation and function delegation—in terms of demands made on human–machine co-operation (HMC) design. Activity allocation is a co-operative activity, as opposed to a private activity. The first section will delineate the framework we have defined in order to analyse co-operation. It will locate activity allocation within the other co-operative activities and introduce a crucial distinction between role, task and function in the context of HMC. The second and third sections will present the respective demands of the two allocation principles on design, illustrating them in the air traffic control domain. In conclusion, we shall develop a number of arguments in favour of function delegation in situations where the machine’s ability to co-operate is restricted.

2. Human–machine co-operation 2.1. A theoretical framework for cognitive co-operation Our approach to co-operation is more process than structure oriented. It describes co-operative activities (and their underlying representations) and must be complemented by other approaches in terms of relations and communication flows between agents. As is the case with Castelfranchi’s (1998) theory, ours puts at its core the notion of (negative or positive) interference (dependence or inter-dependence between different agents’ goals and sub-goals). We consider co-operative activities as being mainly motivated by the management of such interference, in order to facilitate individual tasks or the overall task (Hoc 2001). We define three levels of co-operative activity in relation to the temporal span (or horizon) covered (figure 1). At the action level, co-operation consists of local interference management. At the planning level, it enables the agents to maintain and/or elaborate a COFOR (common frame of reference), a concept similar to common ground (Clark 1996) or a shared mental model (Cannon-Bowers et al. 1993). It is concerned with the representation of the environment as well as the representation of the team’s activity, and it is an internal representation as opposed to an external support (Jones and Jasek 1997). Activity allocation belongs to the planning level. The meta-co-operation level integrates long-term constructs, including ‘translators’ between each partner’s representations or models of oneself, or of the other agents. The activities that can be found at each

Task and function allocation

285

Figure 1. Co-operative activities (after Hoc 2001).

level enable the agent to introduce an anticipative feature into the activities situated at the previous level. For example, a model of the other agent can facilitate goal identification. Diverse forms of interference are possible. Some of them have already been widely studied by artificial intelligence and psychology in the domain of planning (Hoc 1988)—precondition or interaction between goals. Others are specific to co-operation (Hoc 2001)—redundancy between agents or mutual control (of each agent over the others’ activities). Redundancy is obviously a necessary condition for activity allocation. As far as HMC is concerned, mutual control is crucial since the machine is always supervised by a human operator and can therefore contribute to the correction of human errors. 2.2. The problem of activity allocation between humans and machines Within the HMC context, the use of concepts borrowed from human–human co-operation is justified because it has been proved that humans can transfer co-operative attitudes to machines (Nass et al. 1996). However, to use concepts originating from the study of human–human co-operation, without a degree of caution, in order to approach human–machine relationships could be inappropriate. The most debatable is certainly the notion of role in which there are two components: the activity and the related responsibility. Role allocation is possible between humans, but one cannot allocate a ‘role’ to a machine that is only able to assume a certain authority, rather than any responsibility. If an activity is allocated to a machine, there is always a ‘front-line’ human operator responsible for this activity. One particularly well-documented difficulty with automation is the complacency phenomenon whereby human responsibility is reduced without replacing it with any ‘machine’ responsibility (Hoc 2000). Certainly, in law, there is always a

286

J.-M. Hoc and S. Debernard

responsible entity, but it is always of a human nature (the human manager, the human designer or the front-line human operator). Strictly speaking, there is no role allocation between humans and machines. The notion of task allocation looks more realistic than role allocation in HMC, that is to say, it is a goal that can be reached with some autonomy. The designers have no problem in performing hierarchical task analyses, decomposing overall tasks into subtasks, sub-subtasks, and so on, until they reach elementary levels. However, is putting one foot in front of the other a task for a walker under ordinary conditions? When we are confronted by a human operator performing what we think of as a task, several problems can arise. An operation such as taking a step may be considered by the walker to be simply a means rather than an end. Elaborating a representation of this particular goal can jeopardize the necessary fluidity of the walk. Furthermore, when the walker stoops under a heavy burden, the high level of interaction between the two operations makes it difficult for them to be processed independently. This task decomposition is not then appropriate for the actual execution of this activity. We have defended elsewhere (Hoc 1988) the importance of the consideration of the subject’s task representation to an understanding of task execution. During its execution, a task (or subtask) can be identified by the representation of a goal as an intention to be protected (by the performer). Dynamic task allocation between a human operator and a machine (or another human operator) is acceptable only if the designer’s overall task decomposition is compatible with the task decomposition considered by the operator. A weak form of task allocation is function allocation, where human responsibility for the overall task is recognized. A particular function may be considered, sometimes as a task, sometimes as a means without any goal in itself, but just related to a superordinate goal. A function is more generic than a task because it can be utilized in the performance of different tasks. The same set of functions can be considered within different decompositions of the same (overall) task. However, dynamic function allocation is not acceptable if the human operator cannot identify the function to be allocated. Below, we will discuss function delegation as a particular function allocation mode where the human operator decides to allocate a function to a machine explicitly, within the framework of a task decomposition given by the human. 2.3. An example: air traffic control In France, a research programme has been developed to explore the possible benefits of adaptive automation as a solution to the yearly 7% increase in air traffic in Europe (Vanderhaegen et al. 1994, Hoc and Lemoine 1998, Debernard et al. 2002). It has made use of diverse automatic conflict resolution devices (ACRDs) to allocate dynamically some activities, either to the human controller or to the ACRD, in order to alleviate the human workload. The distinction we have just made between task and function applies clearly to this programme. Figure 2 reproduces part of an air traffic control radar-like scope. At first, two aircraft (AFR124 and BAW456) can be identified as conflicting (crossing on CGC beacon under the acceptable separation gap) if nothing is done (step 1). The third one (BAW678) is not conflicting. Following the usual rule (that governs relations between aircraft speeds), the (human) air traffic controller decides to make AFR124 go behind BAW456 (step 2). As soon as this intention is formulated in his/her mind, the contextual aircraft BAW678 ‘enters’ into the problem, conflicting with AFR124. A second deviation decision is made (step 3) and the problem is completely resolved (step 4). This presentation of the problem is very analytical as opposed to the representation likely to be elaborated by human controllers who use powerful pattern recognition

Task and function allocation

287

Figure 2. Resolution of a three-aircraft conflict in air traffic control. Step 1: three aircraft are going to CGC beacon. A two-aircraft conflict appears between AFR124 and BAW456. Step 2: if AFR goes behind BAW456, BAW678 becomes conflicting with AFR. Steps 3 and 4: a second heading instruction must be implemented to resolve the entire problem.

processes (Klein et al. 1993). As a matter of fact, controllers recognize a three-aircraft conflict immediately and the two-step solution is envisioned straightaway. The first ACRD utilized in the experiments was a two-aircraft conflict resolution device. The task unit represented by the controllers is not of this kind. The two successive twoaircraft conflicts are not resolved separately, but as a means (function applications) of executing the three-aircraft task. In addition, if the device turns AFR124 to the left in the two-aircraft conflict (AFR124 and BAW456), it can create a more complex problem, involving several aircraft on the left, than the initial three-aircraft problem. This analysis also shows that the way a task is defined is not independent of intention and that task decomposition is governed by the reduction of interference between subtasks.

288

J.-M. Hoc and S. Debernard

3. Dynamic task allocation 3.1. Task definition Dynamic task allocation (DTA) assumes that the tasks (and subtasks) to be allocated, either to the human or to the machine, can be defined beforehand in a generic way. This generic definition will then be applied in real time to identify subtasks of this kind, resulting in a decomposition of the overall task. More often than not, the definition relies on competency of the automatic device, since it is crucial to ensure that it can take the place of the human (redundancy). The first part of the air traffic control (ATC) research programme (SPECTRA) explored the concept of DTA. It assumed that ATC controllers could accept DTA when two-aircraft tasks are actually considered (Vanderhaegen et al. 1994, Hoc and Lemoine 1998). It was found that the benefits of using the ACRD were considerably reduced, because of a number of refusals by controllers to allocate (or to see allocated) two-aircraft conflicts belonging to three- or fouraircraft problems. It was felt that the ACRD would render the problem more complex to solve. Despite this, however, the ACRD still appeared to be effective.

3.2. Implicit and explicit allocation A first version of the platform (simulator and assistance: SPECTRA V1 (Vanderhaegen et al. 1994)) was developed to compare two allocation modes: implicit and explicit. Only radar controllers (in charge of safety and expedition in the sector) were employed for night traffic duty. Planning controllers, in charge of regulating the radar controllers’ workload and inter-sector co-ordination, were not present. The implicit mode consisted of imposing the allocation on the basis of an evaluation of the radar controller’s workload and the maintenance of this workload below a certain level. In the explicit mode, the radar controller decided on the allocation. These two modes were compared with a control situation where no ACRD was used. Despite using an inappropriate experimental design (one that lacked balancing order effects), a high degree of consistency was found between different kinds of variables (objective and subjective measures). This led the researchers to draw three main conclusions.  The two modes led to a better performance (e.g. near-misses, fuel consumption, etc.) than the control situation, in correlation with the number of conflicts allocated to the ACRD.  The implicit mode was better than the explicit mode, in terms of performance, for two reasons. Firstly, the explicit mode mixed strategic (allocation) and tactical (conflict resolution) activities, which led to an overload. It has already been shown that, in aircraft piloting, the captain is in charge of strategic activities and, when flying, the addition of this tactical activity can lead to incidents (Jentsch et al. 1999). Secondly, implicit task allocation relied on a two-aircraft conflict definition that, sometimes, was not compatible with the traffic decomposition made by the controllers. In the explicit mode, such conflicts were not allocated to the ACRD to avoid negative interference with the controller’s activity. This resulted in fewer conflicts being allocated to the ACRD in the explicit mode (and more workload) than in the implicit mode.  Despite its positive effect on performance, however, the implicit mode was less appreciated by the controllers than the explicit mode. They reported that they were very anxious to keep control over the entire situation.

Task and function allocation

289

A second platform (SPECTRA V2 (Hoc and Lemoine 1998)) was developed in order to reduce the explicit allocation demands. This necessitated the use of two kinds of controller—radar controllers and planning controllers. Two explicit modes were defined—a fully explicit mode and an assisted explicit mode. The fully explicit mode enabled the two controllers to allocate the conflicts. In the assisted explicit mode, the machine made a proposal, the planning controller could exert a veto right, but the radar controller was not in charge of the allocation. The two modes were compared with a control situation and the order effects were balanced. The main results were as follows (obviously, the problem of task decomposition remained).  The effect of the two allocation modes was positive as compared to the control situation, not only in terms of performance, but also in terms of private and co-operative activities. Private strategies were more anticipative (conflicts detected earlier, resolved by fewer manœuvres, activities determined more by early intentions than by responses to alarms or to salient information, etc.) because time was saved by the ACRD resolutions (including rerouting) and the planning controllers’ contribution to DTA. Co-operative activities (between human controllers) appeared to be easier to develop (more efficient communication), possibly because of a richer external COFOR on the interfaces.  The assisted explicit mode was more effective than the fully explicit mode, in terms of performance and co-operation.  However, some evidence of a complacency phenomenon was found in the assisted explicit mode protocols in comparison with the fully explicit mode protocols. The radar controllers could not allocate conflict in the assisted explicit mode and thus felt themselves to be less responsible for the tasks performed by the machine. On the contrary, in the fully explicit modes they were much more critical of the ACRD solutions. Complacency was very often described in the human–machine relation (Hoc 2001). It is not an excess of trust (overconfidence) in the machine, just a lack of mutual control of the human over the machine (for workload or motivation reasons) and a disjunction of the supervision fields between human and machine action areas.

3.3. The demands of dynamic task allocation on design From these SPECTRA experiments, four conclusions can be drawn when it comes to designing an efficient DTA between a human and an artificial agent.  There must be compatibility between task decompositions. The best results ought to be attained by decomposition into almost independent subtasks, considering the human intentions. If the machine is unable to produce such an almost independence and to infer intentions, DTA cannot be entirely satisfactory.  Human control should be maintained over the situation. Explicit DTA is always the best and purely implicit DTA should be avoided. A compromise between the two modes is the production of proposals, which are then validated. However, the machine should be able to produce acceptable proposals on a frequent basis.  Human responsibility should be retained within the situation. Human operators should commit themselves in the allocation process in order to avoid complacency, that is to say a split in the supervision field, where only they supervise their own action fields. Consequently, implicit DTA should be avoided.

Q1

290

J.-M. Hoc and S. Debernard

 A mixture between strategic and tactical tasks should be avoided. There is a compromise to manage between the human implication in the allocation and the possible overload.

4. Dynamic function delegation 4.1. Task, intention and function definition In dynamic situations, tasks and intentions are defined in real time. Dynamic function delegation (DFD) implies that the human operator defines the tasks. The second step of the ATC programme (AMANDA: Debernard et al. 2002) has led to the design of a new platform where the two controllers are in charge of defining tasks. The new ACRD (called STAR, the French acronym for resolution assistance system) can perform various functions, interacting with the controllers. The diverse forms of interaction contribute to define a human–machine interface (HMI) to be used as a support to co-operation between STAR and the controllers and between the controllers themselves. The HMI corresponds to an external COFOR called Common Work Space (CWS) (see Bentley et al. 1992, Decortis and Pavard 1994, Jones and Jasek 1997). The CWS has been derived from Rasmussen’s (1983) model, revised by Hoc and Amalberti (1999) and contains several ‘attributes’ (figure 3).     

Information stemming from information elaboration activities. Problems stemming from identification activities. Strategies stemming from schematic decision-making activities. Solutions stemming from precise decision-making activities. Instructions stemming from solution implementation activities.

Figure 3. Common Work Space (CWS) and co-operative activities.

Task and function allocation

291

To co-operate, the human agents and the assistance tool are supposed to pass on information, problems, strategies, etc. in order to share their own frame of reference. Then, on the basis of the CWS, it is possible to use the three co-operation forms defined by Schmidt (1991) to characterize the co-operative activities (Pacaux-Lemoine and Debernard 2002).  In the debative form (figure 3(a)), all the agents supply the CWS with new data (for a function). When some interference appears, they can negotiate.  In the integrative form (figure 3(b)), only one agent supplies the CWS with data. The other agent uses the data to perform the next function.  In the augmentative form (figure 3(c)), the agents perform the same type of function and update the CWS, but for different function applications, in accordance with the allocation of these applications. The implementation of a CWS between an assistance tool and one or several human agents introduces some constraints, especially the need for negotiation. Two human agents, when they negotiate, may use schematic explanations that are very efficient. An artificial agent needs a precise explanation on the basis of task-related information. For the moment, it is very difficult, in an artificial agent, to implement human negotiation capabilities. In the case of AMANDA, the CWS’s attributes have been extracted from a cognitive analysis of human co-operative activities in an experiment where two radar controllers were sharing the same traffic (Hoc and Carlier 2002) (figure 4). The tasks (problems to be resolved) are defined by the controllers as ‘clusters’ of related aircraft. STAR does not intervene without any imposed task definition (e.g. the cluster in figure 5). Within a task, a strategy is defined as one or several under-specified plans to be fully specified, for example, ‘make EIN566 go behind KLM1707’ (and

Figure 4. The AMANDA’s CWS, the STAR’s functions and interaction with controllers.

292

J.-M. Hoc and S. Debernard

Figure 5. The AMANDA resolution display. The controller has built a three-aircraft cluster (EIN566, KLM1707 and UPS6356). The initial conflict is composed of EIN and KLM on VELIN beacon. Due to the EIN’s speed, the controller has entered the directive ‘UPS goes behind KLM’ (‘passer derrie`re’). STAR proposes a solution to avoid this conflict, but detects a new conflict between EIN and UPS. The controller enters the new directive ‘UPS goes behind EIN’ and STAR finds a new plan that resolves the problem. The display was printed when the two directives had been entered and one of them implemented (heading instruction sent to UPS).

‘reroute it as soon as possible’ is always implied by any plan) (see figure 5). These underspecified plans are called ‘directives’. The machine is then used as a super-calculator to compute an acceptable deviation (heading or level instruction) that is not implemented (sent to the aircraft) at once (delayed instruction). If necessary, the controllers can introduce further constraints, such as ‘go behind X after a certain point’. If they have not noticed the problem with the third aircraft (UPS6356), the ACRD tells them that the plan is not feasible because of this other conflict. Then, the controllers can add the second plan—‘make UPS6356 go behind EIN566’—and receive a final validation of the solution. DFD takes place when the controller orders plan execution. 4.2. The demands of dynamic function delegation on design This new principle enables a quite simple machine to co-operate with the controllers without introducing negative interference (destroying a correct task representation, producing new problems to solve in the future, etc.). It also enables the machine to partici-

Task and function allocation

293

pate in the mutual control activity (improving task representation). Delegating a function is not a strategic activity since it is fully integrated into the tactical conflict resolution activity. An evaluation of this principle is in progress in order to validate the balance between the development of costly attentional activities inside what could be routinized activities and the benefits of delegation. In comparison with DTA, DFD is less demanding on machine design since there is no need for task decomposition and no incompatibility problems. Human control and responsibility over the situation are protected. Demands on design are restricted to direct manipulations of interfaces fully compatible with radar scopes in order to avoid destroying those routinized activities that are absolutely necessary in ATC.

5. Conclusion Dynamic activity allocation and adaptive automation are both reactions to numerous criticisms of full automation and illustrate its drawbacks (Parasuraman 1997, Hoc 2000), including: a decrease in the HMS adaptive power; an increase in risks because of human complacency about a badly designed machine; and so on. Certainly, there should be continued research into the design of machines with more know-how and more ability to co-operate. Complex task decomposition, intention recognition and co-operative planning should also be greatly improved. However, with the present state of affairs, DTA remains difficult to accept in real dynamic situations where human expertise should be promoted rather than impoverished or badly advised by narrow-minded machines. That is why, after having explored DTA in ATC, we have adopted the best benefit of DFD. Now, returning to the main topics of this special issue, our experience suggests the following.  The human agent should decide the allocation (possibly with computer support).  Intention recognition is needed in task definition, but human intentions can be transmitted to the machine at low cost (schematic plans).  Monitoring (mutual control) can be promoted in both directions, in other words, not only from the human to the machine, but also in the reverse direction.  Shared knowledge and plan (COFOR) are necessary. As is function delegation within common problem representations where feasible at a low cost.  The machine’s autonomy must be restricted if it is likely to produce negative interference in human activity.

Acknowledgement This research programme has been funded by the French National Research Center on Air Traffic Management (CENA).

References Bainbridge, L., 1987, Ironies of automation. In J. Rasmussen, K. D. Duncan and J. Leplat (eds) New Technology and Human Error (Chichester: Wiley), pp. 271–284. Bentley, R., Rodden, T., Sawyer, P., and Sommerville, I., 1992, An architecture for tailoring cooperative muti-user display. In Proceedings of CSCW’92 (New York: ACM Press), pp. 187–194. Cannon-Bowers, J. A., Salas, E., and Converse, S. A., 1993, Shared mental model in expert decision making teams. In N. J. Castellan, Jr. (ed.) Current Issues in Individual and Group Decision Making (Hillsdale NJ: Lawrence Erlbaum Associates), pp. 221–246.

Q2

294

J.-M. Hoc and S. Debernard

Castelfranchi, C., 1998, Modelling social action for agents. Artificial Intelligence, 103: 157–182. Clark, H. H., 1996, Using Language (Cambridge: Cambridge University Press). Debernard, S., Cathelain, S., Crevits, I., and Poulain, T., 2002, AMANDA project: delegation of tasks in the air traffic control domain. In M. Blay-Fornarino, A. M. Pinna-Dery, K. Schmidt and P. Zarate´ (eds) Cooperative Systems Design (Amsterdam: IOS Press), pp. 173–190. Decortis, F., and Pavard, B., 1994, Communication et coope´ ration: de la the´ orie des actes de langage a` l’approche ethnome´ thodologique [Communication and cooperation: from language acts theory to ethnomethodological approach]. In B. Pavard (ed.) Syste`mes Coope´ ratifs: De la Mode´ lisation a` la Conception (Toulouse: Octare`s), pp. 21–50. Fitts, P. M., 1951, Human Engineering for an Effective Air Navigation and Traffic Control System (Washington, DC: National Research Council). Hoc, J. M., 1988, Cognitive Psychology of Planning (London: Academic Press). Hoc, J. M., 2000, From human–machine interaction to human–machine cooperation. Ergonomics, 43: 833–843. Hoc, J. M., 2001, Towards a cognitive approach to human–machine cooperation in dynamic situations. International Journal of Human–Computer Studies, 54: 509–540. Hoc, J. M., and Amalberti, R., 1999, Analyse des activite´ s cognitives en situation dynamique: d’un cadre the´ orique a` une me´ thode [Cognitive activity analysis: from theoretical framework to method]. Le Travail Humain, 62: 97–130. Hoc, J. M., and Carlier, X., 2002, Role of a common frame of reference in cognitive cooperation: sharing tasks between agents in air traffic control. Cognition, Work, & Technology, 4: 37–47. Hoc, J. M., and Lemoine, M. P., 1998, Cognitive evaluation of human–human and human–machine cooperation modes in air traffic control. International Journal of Aviation Psychology, 8: 1–32. Hollnagel, E., and Woods, D. D., 1983, Cognitive systems engineering: new wine in new bottles. International Journal of Man–Machine Studies, 18: 583–600. Jentsch, F., Barnett, J., Bowers, C., and Salas, E., 1999, Who is flying this plane anyway? What mishaps tell us about crew member role assignment and air crew situation awareness. Human Factors, 41: 1–14. Jones, P. M., and Jasek, C. A., 1997, Intelligent support for activity management (ISAM): an architecture to support distributed supervisory control. IEEE Transactions on Systems, Man, and Cybernetics—Part A: Systems and Humans, 27: 274–288. Kaber, D. B., Riley, J. M., Tan, K. W., and Endsley, M., 2001, On the design of adaptive automation for complex systems. International Journal of Cognitive Ergonomics, 5: 37–57. Klein, G. A., Orasanu, J., Calderwood, R., and Zsambok, C. E., 1993, Decision-Making in Action: Models and Method (Norwood NJ: Ablex). McCarthy, J. C., Fallon, E., and Bannon, L., 2000, Dialogues on function allocation. International Journal of Human–Computer Studies, 52 (2). Millot, P., and Mandiau, R., 1995, Man–machine cooperative organizations: formal and pragmatic implementation methods. In J. M. Hoc, P. C. Cacciabue and E. Hollnagel (eds) Expertise and Technology: Cognition & Human–Computer Cooperation (Hillsdale NJ: Lawrence Erlbaum Associates) pp. 213–228. Nass, C., Fogg, B. J., and Moon, Y., 1996, Can computers be teammates? International Journal of Human–Computer Studies, 45: 669–678. Older, M. T., Waterson, P. E., and Clegg, C. W., 1997, A critical assessment of task allocation methods and their applicability. Ergonomics, 40: 151–171. Pacaux-Lemoine, M. P., and Debernard, S., 2002, Common work space for human–machine cooperation in air traffic control. Control Engineering Practice, 10: 571–576. Parasuraman, R., 1997, Humans and automation: Use, misuse, disuse, abuse. Human Factors, 39: 230–253. Rasmussen, J., 1983, Skill, rules, and knowledge: signals, signs and symbols, and other distinctions in human performance models. IEEE Transactions on Systems, Man, and Cybernetics, 15: 257– 266. Rieger, C. A., and Greenstein, J. S., 1982, The allocation of tasks between the human and computer in automated systems. In Proceedings of the IEEE 1982 International Conference on ‘Cybernetics and Society’ (New York: IEEE), pp. 204–208.

Q3

Task and function allocation

295

Schmidt, K., 1991, Cooperative work: a conceptual framework. In J. Rasmussen, B. Brehmer and J. Leplat (eds) Distributed Decision-Making: Cognitive Models for Cooperative Work (Chichester: Wiley), pp. 75–110. Vanderhaegen, F., Crevits, I., Debernard, S., and Millot, P., 1994, Human–machine cooperation: toward and activity regulation assistance for different air-traffic control levels. International Journal of Human–Computer Interaction, 6: 65–104.