Automation and Adaptive Behavior

Introduction

In his target article on human–automation interaction (HAI), David Kaber (2018 [this issue]) set himself the task of taking stock of the large and rapidly expanding cognitive engineering literature in this area and distilling what he believes to be the major factors holding back further progress in the modeling, analysis, and design of HAI systems. These factors are too numerous and diverse to serve as the subject of a brief commentary, so I have chosen to focus on what I take to be one of his most critical and incisive observations. In particular, Kaber stresses the point that much, if not most, current HAI research insufficiently finds a place for representing how satisficing behavior on the part of humans plays a role in determining the eventual success or the performance levels that will be achieved in HAI systems of any particular design.

I have chosen to focus on this theme for two reasons. First, I believe it to be one of the most important observations presented in Kaber’s paper. Second, it is one of the many factors discussed that I may be qualified to discuss and elaborate, given that, in one form or another, satisficing (or adaptive behavior more generally) has played a key role in my own research and my research done with my students and collaborators.

In the following, I begin with a brief overview of the satisficing concept and provide my rationale for why I believe it will be useful to build on Kaber’s point by further elaborating this discussion to include adaptive cognition and behavior more broadly, seeing satisficing as just one of many manifestations of the latter. Then, due to space constraints, I have framed my discussion around just three publications. The first of these (Kirlik, 1993) was my first journal article on HAI and, from the best I can tell from Google Scholar, the earliest such publication using the term human–automation interaction in its title. This paper illustrates that automation designers can fail to anticipate one form of adaptive behavior known as satisficing (i.e., having a faulty presumptive model in Kaber’s terms), resulting in their systems going unused. The next (Kirlik, 2006a) was an edited research volume collecting what I judged at the time to be a critical mass of research, much of it related to HAI, and whose title began with the phrase “Adaptive Perspectives.” This volume is devoted to theoretical and methodological techniques for the study of adaptive behavior generally speaking, going beyond Kaber’s call for more research on satisficing to also include adaptive cognition and behavior in inference and situation awareness. Finally, I conclude by briefly discussing our most recent research project on HAI to date, as reported in a journal article (Ackerman et al., 2017) in which the goal was to repair breakdowns in HAI (in aviation) due to the fact that many current cockpit interfaces do not provide sufficient information to allow adaptive pilot cognition and behavior to flourish rather than flounder. This paper illustrates not only the benefits of providing information sources allowing adaptive behavior to flourish but also the difficulty of predicting the nature of highly adaptive, expert levels of performance.

Satisficing and Adaptive Behavior

Kaber correctly gives credit to Simon (e.g., Simon, 1957; see also Kirlik & Bertel, 2010) for putting forth the idea that humans (as well as other animals) often simplify decision making by selecting the first option that exceeds some specified aspiration level (satisfice), as opposed to engaging in the more cognitively demanding method of optimizing (cf. Canellas & Feigh, 2016). For Kaber, this observation serves as justification for many of his arguments: If HAI designers assume that humans will use automation in an optimal way (most frequently, not in the sense of true mathematical or utility-theoretic optimality but instead akin to “as the designer expects or envisions them to do”), then we can expect mismatches between HAI performance as intended and HAI performance as observed.

I believe, however, that it will be useful to expand the scope of the discussion around satisficing to include the more general concept of adaptive cognition and behavior. The term satisficing is an entrenched concept in both the behavioral decision-making and decision-science communities, associated with the selection of an action from a set of alternatives. Yet, something akin to satisficing exists on the input side of cognition as well—in making judgments or predictions about the system or environment or, more generally, in achieving situation awareness (Endsley, 1995; see also Kirlik & Strauss, 2006).

Researchers focusing on these abilities, such as those studying fast and frugal heuristics (Gigerenzer & Selten, 2001; see also Bertel & Kirlik, 2011) have found that, in addition to satisficing in making decisions, humans are also highly adaptive when it comes to making judgments and predictions. Here, the analogy to satisficing involves using the most readily available, salient, and minimal set of cues or information sources available, often with surprising accuracy (e.g., as illustrated in heuristics such as “take the first,” “take the best,” or simple yet robust linear additive cue-combination rules; see Gigerenzer & Brighton, 2009, for arguments on why simple, inferential rules often outperform more complex rules, especially in uncertain tasks).

If humans do use heuristics such as these as a form of adaptive cognition and behavior in HAI and thus rely on more restricted sets of available information than HAI designers intend them to do, we can once again expect to see mismatches between HAI performance as intended and HAI performance as observed.

Modeling Strategic Behavior in HAI: Why an Aid Goes Unused

The article discussed here emerged from my dissertation research (Kirlik, Miller, & Jagacinski, 1993) and is a direct illustration of Kaber’s observation about the human tendency to satisfice in decision making or strategic behavior. This article describes a computational model of human supervisory control performance grounded in J. J. Gibson’s (1979) theory of behavior as an adaptive reflection of the affordance structure of the environment. Yet, that research also resulted in a second publication focused on one of the more surprising results from our human-in-the-loop experimentation: Participants in our studies never used an autopilot in the way we (the researchers and designers) intended, that is, as a task off-load aid when exogenous events required them to divert their attention from their primary task of piloting a simulated helicopter.

This finding motivated me to determine why, and the resulting article (Kirlik, 1993), using Markov decision process modeling and dynamic programming, provided a startling answer: Given the automation design with which they were confronted, it would not have been adaptive or rational for our participants to use automation in the manner intended—a clear example of a mismatch between expected and observed HAI performance. Had we, as automation designers, performed the type of modeling and sensitivity analysis presented in that paper to address the question, “How should we design an autopilot so that adaptive performers would embrace its use and achieved its benefits?” prior to its design, it is unlikely that we would have observed such mismatches between what was intended and observed in HAI performance.

Adaptive Perspectives in Human–Technology Interaction

This publication (Kirlik, 2006a) was an edited research volume in which the first sentence of the overview of the book stated, “An emerging trend in cognitive science and psychology is to investigate cognition as constrained adaptation to the environment, and today, adaptation often means working with or through technology” (p. n.p.). For current purposes, this volume is important in that it illustrates the value of viewing satisficing as just one instance of the more general phenomenon of adaptive behavior. Although some of the chapters in the volume were focused on research related to interface design and to technological systems other than those of central interest to cognitive engineering (e.g., the World Wide Web), the majority of chapters in the volume provided detailed studies in which modeling was used to help describe and understand adaptive HAI system performance.

Although this volume has had impact in some quarters, it has had less impact on cognitive engineering, human factors, and human–computer interaction research practice than I had originally hoped. I suspect at least some of this was due to a less than ideal, adaptive mesh between the theoretical and methodological orientations of its contents and those of the lion’s share of HAI research covered by Kaber in his review. Much of the latter is grounded in the dominance of orthogonal thinking in the social and behavioral sciences, as reflected in preferences for systematic rather than representative experimental designs, the influence of the general linear model on how research is conceived, and its associated data analysis techniques, such as ANOVA and its variations (for more on these issues see Hammond, 2006).

To some extent, I believe that factors such as these have contributed to the gap in the literature noted by Kaber as to the paucity of research addressing satisficing and other forms of adaptive cognition and behavior in HAI research.

Promoting Automation Awareness Through Visualization

The final publication to be discussed (Ackerman et al., 2017) focuses on the fact that we cannot expect humans to display successful levels of adaptive behavior in HAI systems if interface designs do not provide the appropriate information for them to do so. This article notes that the flight envelope protection (FEP) automation present in many modern aircraft behaves in ways that are often opaque to the pilot or flight crew. Like the volume discussed previously, this paper also illustrates the importance of extending Kaber’s call for more research on satisficing to include research on adaptive behavior in both decision-making and inferential processes.

In this research, we created and evaluated techniques for visualizing what FEP automation knows (or at least believes) about the relationship between dynamic aircraft states and flight safety envelopes and superimposed this information on the primary flight display or PFD. Additionally, we developed a second, accompanying display that made visible to pilots if and when FEP automation was providing control compensation, and to what extent it was doing so.

A brief public demo of this prototype HAI system in operation is available through FEP Demo (2017). Although this demo should be self-explanatory, note that in the detailed empirical evaluations presented in Ackerman et al. (2017), we explicitly noted that our findings indicated that the design of our system may have underestimated the extremely high levels of adaptive control skill of truly expert pilots, as opposed to the levels of skill exhibited by the flight instructors and pilots with less experience, whose unaided performance motivated our novel designs: “Future research will explore the possibility of using the developed protection and interface system as an upset recovery training aid and for pilot education regarding the safe operation limits of an unfamiliar aircraft” (p. 14).

Conclusion

I fully agree with Kaber’s observation that satisficing and other forms of adaptive cognition and behavior have been insufficiently represented in HAI research. More research is needed on how users of automated systems will seek to simplify decision making by finding adaptive, heuristic solutions to what appear to designers to be complex control tasks and by finding adaptive, heuristic solutions to what appear to designers to be complex judgment or situation awareness tasks. A detailed set of six concrete prescriptions for performing such research is presented in Kirlik (2006b).