Portal de Periódicos da CAPES

Modeling Decision Making on the Use of Automation

2011; Wiley; Volume: 33; Issue: 33 Linguagem: Inglês

1551-6709

Junya Morita, Kazuhisa Miwa, Akihiro Maehigashi, Hitoshi Terai, Kazuaki Kojima, Frank E. Ritter,

Cognitive Science and Mapping

Modeling Decision Making on the Use of Automation Junya Morita (j-morita@jaist.ac.jp) School of Knowledge Science, Japan Advanced Institute of Science and Technology, Japan Kazuhisa Miwa (miwa@is.nagoya-u.ac.jp) Akihiro Maehigashi (mhigashi@cog.human.nagoya-u.ac.jp), Hitoshi Terai (terai@is.nagoya-u.ac.jp) Graduate School of Information Science, Nagoya University, Japan Kazuaki Kojima (koj@aoni.waseda.jp) Faculty of Human Sciences, Waseda University, Japan Frank E. Ritter (frank.ritter@psu.edu) College of Information Sciences and Technology, Penn State, USA Abstract This paper presents a cognitive model that simulates reliance on automation using a line-tracing task similar to driving where an operator has to track a moving line with a circle by pressing keys on a keyboard (manual control) or rely on automation (auto control). An operator can switch between auto and manual control during the task. The success proba- bilities of each control mode were systematically varied. An ACT-R model to perform this task was constructed by repre- senting reliance on the automation as production. The model performs this task through productions that manage the per- ceptual/motor modules. The utility values of these productions are updated based on the rewards in every screen update. We also introduce a meta-level monitoring the internal state of the model. A preliminary run of this model simulated the overall trends of the behavioral data, suggesting some validity of the assumptions made in our model. Keywords: Automation; ACT-R; Trust. Introduction We unconsciously use automation systems like e-mail spam filters, spell checkers, and electronic toll collection (ETC) systems. These systems save time and help us lead more ef- ficient lives. We, however, sometimes face difficult choices about whether to use these systems (Bainbridge, 1983). It has also been pointed out that human decision making using such systems is not always optimal. For example, Parasuraman and Riley (1997) stated that there are two types of maladap- tive choices of using automation: misuse, the over-reliance of automation, and disuse, the underutilization of automation. Some studies indicated that human users have an automa- tion bias towards misuse of automation systems (Bahner, Hu- per, & Manzey, 2008; Singh, Molloy, & Parasuraman, 1997; Skita, Moiser, & Burdick, 2000). On the other hand, other research indicated that human users have a manual bias, a bias towards disuse of automation systems consistent with a need for control (Beck, McKinney, Dzindolet, & Pierce, 2009; Dzindolet, Peterson, Pomranky, Pierce, & Beck, 2003; Dzindolet, Pierce, Beck, & Dawe, 2002). Vries, Midden, and Bouwhuis (2003) experimentally re- vealed that the reliance of automation is influenced by both the Capability of Manual control (Cm) and the Capability of Auto controls (Ca). To explain these effects, Gao and Lee (2006) proposed the Extended Decision Field Theory (EDFT model; Figure 1). The model constructs belief of Ca and Cm (Bca, Bcm) based on partially displayed these values. From the belief values, trust (T) and self-confidence (SC) are con- structed. Preference of automation (P) is determined by sub- tracting T from SC. If P exceeds an upper threshold (θ), then the model turns the current control mode to auto. If P falls below a lower threshold (-θ), then the model turns the cur- rent control mode to manual. In every cycle, values of Bca, Bcm, T, and SC are updated by differential equations. Al- though this model clearly explains the reliance on automation in dynamic situations, the model does not have any knowl- edge about tasks. It cannot interact with a task environment, and it provides no human performance predictions. This report describes a cognitive process model that in- teracts with a specific task environment where sequential decision-making is made. Especially, we extend our previ- ous model (Morita et al., in-press) to improve its motor con- trol. To do this, we use ACT-R, a unified theory of cognition (Anderson, 2007). The following subsection briefly shows features of this architecture relating to our model. ACT-R One of the most important assumptions of ACT-R is modular- ity of cognition. ACT-R is composed of several independent modules: goal, production, declarative, perceptual, and mo- tor (Anderson, 2007). A goal module holds the current task goal and other task related information. A production mod- ule and a declarative module hold procedural and declarative knowledge respectively. Perceptual modules include a vision and an audio module, which take information from an exter- nal environment. A motor module manipulates devices like a keyboard or a mouse in an external environment. Modules other than the production module have buffers to hold tem- porarily information called a chunk. A production module integrates the other modules by production rules, which con- sists of a condition/action pair that is used in sequence with other productions to perform a task. Conditions and actions in production rules are specified with buffer contents of each module.