A Model for Types and Levels of Human Interaction with Automation

Source: Raja Parasuraman, Thomas B. Sheridan, Christopher D. Wickens; IEEE Transactions on Systems, Man, and Cybernetics Part A 30(3), journal article · 2000

Abstract—Technical developments in computer hardware and software now make it possible to introduce automation into virtually all aspects of human-machine systems. Given these technical capabilities, which system functions should be automated and to what extent? We outline a model for types and levels of automation that provides a framework and an objective basis for making such choices. Appropriate selection is important because automation does not merely supplant but changes human activity and can impose new coordination demands on the human operator. We propose that automation can be applied to four broad classes of functions: 1) information acquisition; 2) information analysis; 3) decision and action selection; and 4) action implementation. Within each of these types, automation can be applied across a continuum of levels from low to high, i.e., from fully manual to fully automatic. A particular system can involve automation of all four types at different levels. The human performance consequences of particular types and levels of automation constitute primary evaluative criteria for automation design using our model. Secondary evaluative criteria include automation reliability and the costs of decision/action consequences, among others. Examples of recommended types and levels of automation are provided to illustrate the application of the model to automation design.

Introduction. I. INTRODUCTION C ONSIDER the following design problem. A human operator of a complex system provided with a large number of dynamic information sources must reach a decision relevant to achieving a system goal efficiently and safely. Examples include an anesthesiologist given various vital signs who must decide whether to increase the dosage of a drug to a patient undergoing surgery; an air defense operator given various sensor readings who has to decide whether to shoot down a potentially hostile enemy aircraft; or a securities analyst given various financial data who must judge whether to buy a large block of stocks. Technical developments in computer hardware and software make it possible to automate many aspects of the system, i.e., to have a computer carry out certain functions that the human operator would normally perform. The automation can differ in type and complexity, from simply organizing the information sources, to integrating them in some summary fashion, to suggesting decision options that best match the incoming information, or even to carry out the necessary action. The system design issue is this: given these technical capabilities, which system functions should be automated and to what extent? These fundamental questions increasingly drive the design of many new systems. In this paper we outline a model of human interaction with automation that provides a framework for answers to these questions. The human performance consequences of specific types and levels of automation constitute the primary evaluative criteria for automation design using the model. Secondary evaluative criteria include automation reliability and the costs of action consequences. (Both these sets of criteria are described more fully later in this paper). Such a combined approach—distinguishing types and levels of automation and applying evaluative criteria—can allow the designer to determine what should be automated in a particular system. Because the impact of the evaluative criteria may differ between systems, the appropriate types and levels of automation for different systems can vary widely. Our model does not therefore prescribe what should and should not be automated in a particular system. Nevertheless, application of the model provides a more complete and objective basis for automation design than do approaches based purely on technological capability or economic considerations.

Machines, especially computers, are now capable of carrying out many functions that at one time could only be performed by humans. Machine execution of such functions—or automation—has also been extended to functions that humans do not wish to perform, or cannot perform as accurately or reliably as machines. Technical issues—how particular functions are automated, and the characteristics of the associated sensors, controls, and software—are major concerns in the development of automated systems. This is perhaps not surprising given the sophistication and ingenuity of design of many such systems (e.g., the automatic landing of a jumbo jet, or the docking of two spacecraft). The economic benefits that automation can provide, or are perceived to offer, also tend to focus public attention on the technical capabilities of automation. In contrast to the voluminous technical literature on automation, there is a small but growing research base examining the human capabilities involved in work with automated systems [1]–[8]. This work has shown clearly that automation does not simply supplant human activity but rather changes it, often in ways unintended and unanticipated by the designers of automation [8], and as a result poses new coordination demands on the human operator [7]. Until recently, however, these findings have not had much visibility or impact in engineering and design circles. Examination of human performance issues is especially important because modern technical capabilities now force system designers to consider some hard choices regarding what to automate and to what extent, given that there is little that cannot be automated. In the present paper we propose a model for types and levels of automation that provides a framework and an objective basis for making such choices. Our approach was guided by the concept of “human-centered automation” [9] and by a previous analysis of automation in air traffic control (ATC) [10].1 Let us begin by defining automation, because the term has been used many different ways. The Oxford English Dictionary (1989) defines automation as 1) Automatic control of the manufacture of a product through a number of successive stages; 2) the application of automatic control to any branch of industry or science; 3) by extension, the use of electronic or mechanical devices to replace human labor. The original use of the term implies automatic control (automatic having many alternative definitions suggesting reflexive action, spontaneity, and independence of outside sources).

Method. In our definition, automation refers to the full or partial replacement of a function previously carried out by the human operator. This implies that automation is not all or none, but can vary across a continuum of levels, from the lowest level of fully manual performance to the highest level of full automation. Several levels between these two extremes have been proposed [11], [12]. Table I shows a 10-point scale, with higher levels representing increased autonomy of computer over human action [10], based on a previously proposed scale [11]. For example, at a low level 2, several options are provided to the human, but the system has no further say in which decision is chosen. At level 4, the computer suggests one decision alternative, but the human retains authority for executing that alternative or choosing another one. At a higher level 6, the system gives the human only a limited time for a veto before carrying out the decision choice. Automated systems can operate at specific levels within this continuum. For example, a conflict detection and resolution system that notifies an air traffic controller of a conflict in the flight paths of two aircraft and suggests a resolution would qualify as level 4 automation. Under level 6 or higher, the system would automatically execute its own resolution advisory, unless the controller intervened. In the proposed model we extend Table I to cover automation of different types of functions in a human-machine system. The scale in Table I refers mainly to automation of decision and action selection, or output functions of a system. However, automation may also be applied to input functions, i.e., to functions that precede decision making and action. In the expansion of the model, we adopt a simple four-stage view of human information processing (see Fig. 1). The first stage refers to the acquisition and registration of multiple sources of information. This stage includes the positioning and orienting of sensory receptors, sensory processing, initial pre-processing of data prior to full perception, and selective attention. The second stage involves conscious perception, and manipulation of processed and retrieved information in working memory [13]. This stage also includes cognitive operations such as rehearsal, integration and inference, but these operations occur prior to the point of decision. The third stage is where decisions are reached based on such cognitive processing. The fourth and final stage involves the implementation of a response or action consistent with the decision choice. This four-stage model is almost certainly a gross simplification of the many components of human information processing as discovered by information processing and cognitive psychologists [14]. The performance of most tasks involves inter-dependent stages that overlap temporally in their processing operations [15]. The stages can also be considered to be coordinated together in “perception-action” cycles [16] rather than in a strict serial sequence from stimulus to response. Our goal is not to debate the theoretical structure of the human cognitive system but to propose a structure that is useful in practice. In this respect, the conceptualization shown in Fig. 1 provides a simple starting point with surprisingly far-reaching implications for automation design. Similar conceptual models have been found to be useful in deriving human factors recommendations for designing systems in general [17]. The four-stage model of human information processing has its equivalent in system functions that can be automated. Accordingly, we propose that automation can be applied to four classes of functions (see also [18] and related proposals in [9] and [19]):

1. information acquisition; 2) information analysis; 3) decision and action selection; 4) action implementation.

Discussion. An important consideration in deciding upon the type and level of automation in any system design is the evaluation of the consequences for human operator performance in the resulting system (i.e., after automation has been implemented). As shown in Fig. 3, particular types and levels of automation are evaluated by examining their associated human performance consequences. To take a hypothetical example, suppose prior research has shown (or modeling predicts) that compared to manual operation, both human and system performance are enhanced by level 4 automation but degraded by automation above level 6. Application of our framework would determine the lower and upper bounds of automation to be 4 and 6, respectively. This initial specification would then be evaluated again with respect to the secondary evaluative criteria, in an iterative manner, and a final choice of level within this range could be made (see Fig. 3). Over the past two decades, researchers have examined a number of different aspects of human interaction with automated systems. This research, which has included theoretical analyzes, laboratory experiments, simulation and modeling, field studies, and analyzes of real-world incidents and accidents, has found that automation can have both beneficial and negative effects on human performance [1]–[10], [45]–[48]. We briefly discuss four human performance areas: mental workload, situation awareness, complacency, and skill degradation. 1) Mental Workload: The evidence suggests that well-designed information automation can change human operator mental workload to a level that is appropriate for the system tasks to be performed. At the simplest level, organizing information sources, e.g., in a priority list, will help the operator in picking the information relevant to a decision. Data summaries can also help by eliminating time-consuming search or communication operations. As mentioned previously, the electronic data block on the air traffic controller’s radar display replaces the need for the controller to communicate with pilots to determine the aircraft position and altitude. Other information automation operations that are beneficial include highlighting, and integration, in which different information sources are collated and presented together [10]. Cockpit predictor displays have also shown that pilot workload decreases and hazard detection performance improves with the addition of predictive information concerning the flight path of neighboring aircraft [21]. Data transformation, for example graphic presentation of information, can also be beneficial. Transformation and integration of raw data into a form (graphical or otherwise) that matches the operator’s representation of system operations has been found to be a useful design principle [49]. A good example is the horizontal situation indicator in the cockpit, which provides the pilot with a graphic display of the projected flight plan and the current position of the aircraft. This, more than any other automated system in the cockpit, has been credited with reducing the workload of the pilot [50]. These results should not be construed to mean that automation always results in balanced operator workload. Instances of automation increasing workload have also been found [8], [50]. These mostly involve systems in which the automation is difficult to initiate and engage, thus increasing both cognitive workload [51] and if extensive data entry is required, the physical workload of the operator. Such systems have been referred to as implementing “clumsy” automation [50]. In general, the effect of automation on mental workload has been mirrored by the similarly mixed record of automation in improving human productivity and efficiency [52].

Conclusion. Automation design is not an exact science. However, neither does it belong in the realm of the creative arts, with successful design dependent upon the vision and brilliance of individual creative designers. (Although such qualities can certainly help the “look and feel” and marketability of the automated system—see [83]). Rather, automation design can be guided by the four-stage model of human-automation interaction we have proposed, along with the consideration of several evaluative criteria. We do not claim that our model offers comprehensive design principles but a simple guide. The model can be used as a starting point for considering what types and levels of automation should be implemented in a particular system. The model also provides a framework within which important issues relevant to automation design may be profitably explored. Ultimately, successful automation design will depend upon the satisfactory resolution of these and other issues.

Limitations. Before concluding, we briefly consider two alternative approaches to the implementation of automation, and discuss some limitations and extensions of our framework. One alternative to our approach is to automate everything that one can. This can be