Modeling Interruption and Resumption in a Smartphone Task: An ACT-R Approach

1.1 Matter of Interruption

According to Brixey et al. (2007), interruption as a certain kind of human experience is usually neither planned nor expected and represents a cognitive break with the task performed up to that time. It can be induced by internal or external sources, resides within a given situational context and indicates a delay in finishing the previous activity. The main goal after facing an interruption comprises successfully returning the mental resources to the actual focus of attention, commonly denoted as resumption. Interruptions are known to impair the main task performance due to a set of disruptive aspects, including those which have a great similarity to the main task (Gillie and Broadbent 1989), an immediate occurrence (Trafton et al. 2003) or the lack of opportunity to refuse or delay the interruption (Salvucci and Taatgen 2010).

Cognitive processes in the face of external interruptions are described by Altmann, Trafton and colleagues (Trafton et al. 2003). They postulate a time course model of interruption and resumption: After starting a main task and performing it for some time, an alert appears announcing the interruption before it actually occurs. The time span between the alert and the upcoming interruption is called interruption lag and is supposed to prepare for effectively returning to the main task. The resumption lag as the interval between ending the interruption and successfully resuming the main task comprises an authentic measure for the extent of disruptiveness. In theoretical accounts, this model is based on the memory-for-goals theory described by Altmann and Trafton (2002), assuming an activation decay of the cognitive representation of the main task in aid of the cognitive representation of the interruption. Such decay might be reduced by either rehearsing information related to the main task or linking environmental cues to certain aspects of the main task.

1.2 Resource Limitation

When dealing with interruptions, one important restricted resource in information processing exists due to working memory limitations, both in terms of duration and capacity (Wickens et al. 2013). The first aspect refers to the fact that information in working memory decays after a certain time. In order to extend such periods of availability, people can rehearse relevant information. In contrast, the matter of capacity indicates that just a defined amount of information can be held active at the same time. According to Miller (1956), it should be between five and nine items, although more recent research proposes smaller numbers (Cowan 2010, Cowan, Morey and Chen In press). Again, rehearsing information depicts a way to increase this span. In general, when performing a memory-related task, memory load has to be maintained by working memory (Anderson, Reder and Lebiere 1996). Increasing such load might affect task performance and result in difficulties to retrieve necessary information.

2.1 Cognitive Modeling Approach

In the field of human-computer interaction, applying cognitive user models is gaining more and more attention. Above all, this method aims to understand task-related human behavior by inspecting underlying cognitive processes. Especially cognitive architectures have proven of value within this context since they provide a theoretical framework to explain basic and constant mechanisms of human cognition behind a variety of tasks (Gray, Young and Kirschenbaum 1997). Moreover, they offer a computational platform for model execution and in this vein give the opportunity to derive predictions on task-related behavior direct.

So far, the Lisp-based cognitive architecture ACT-R (Adaptive Control of Thought – Rational) developed by John R. Anderson and colleagues (Anderson and Lebiere 1998) has been used actively within a vibrant and growing research community to address a variety of basic and applied subjects. It comprises the assumption of different modules that serve as the foundation of any human behavior, occupy defined responsibilities and interact in certain ways to create cognitive processing (Anderson 2007). Figure 1 gives an overview of the standard modules contained in the current ACT-R distribution. These modules have been validated by human experimental data and hold a vested biological background as well since fMRI studies (Anderson 2007, Borst and Anderson 2015) indicate the association of each module with a brain region relevant to the respective cognizance. Each module holds a buffer, serving as an interface to enable communication among the modules. Information is processed within the outlined structure via chunks, i. e. small units encoding relevant elements of knowledge affiliated with a certain category (chunk-type) and containing specific attributes (slots).

Although processes in different modules can be executed in parallel, each buffer can hold just one chunk at the same time, representing the existing limitations in information processing resources. Interaction between modules happens by means of production rules, consisting of a condition and an action part and depicting the main duty of the procedural module. It scans the buffer’s contents and, based on the resulting pattern, selects a suitable production rule that initiates the related action. In case more than one production rule fits, subsymbolic mechanisms apply and a cost-benefit function (utility) decides which production rule is selected. If and how fast a chunk can be retrieved from declarative memory depends on another subsymbolic mechanism called activation. It reflects the availability of information and is determined by the respective context and history of use.

Figure 1

Overview of modules contained in ACT-R 6.0. Adapted from Borst and Anderson (2015) and Anderson (2007).

2.2 Inspected Task Setting

To examine cognitive processes in terms of an interrupted task, a shopping scenario was employed within this project. It operated on a simple smartphone shopping list application (Prezenski and Russwinkel 2014), illustrated in Figure 2, and consisted of encoding, remembering, searching for and selecting a set of 12 predefined products, equally distributed within three runs. Products appeared in a fixed sequence listed on the screen at the beginning of each run. At the end of the task, all products still remembered from the whole selection part had to be recalled.

Figure 2

Main menu, store menu, example product menu for drugstore and example product advertisement for “body lotion” within the shopping list application.

During two of the three runs, interruptions in terms of product advertisements occurred with varying frequency. They were always related to the previously selected shop, announced a special offer, and were triggered by the successful selection of a certain amount of products within the respective run. In order to end the interruption and return to the product selection, a decision for or against the offered product was forced. In conditions with enhanced mental demands, shopping was done for three different people. In consequence, additional information about whom the product should be bought for had to be remembered and recalled throughout the task as well.

Task-related cognition was assessed in terms of certain behavioral performance parameters. Product selection time was computed as the time difference between successfully selecting a product and returning to the related shop menu. For the number of selected products, correctly selected products per run were counted. Resumption time comprised the difference between the offset of the interruption and the transition back to the shop menu. The final recall performance was assessed by summing up correctly recalled products after finishing the selection part.

2.3 Creating the Model

Based on the outlined assumptions from relevant literature on similarity (Gillie and Broadbent 1989), immediacy (Trafton et al. 2003) the forced occurrence of interruptions (Salvucci and Taatgen 2010) and the role of working memory demands (Anderson, Reder and Lebiere 1996, Wickens et al. 2013), an ACT-R model was established. Key features of the described shopping list application were implemented in the ACT-R experimental GUI in a simplified way. In contrast to the model of Trafton et al. (2003), no alert was included, resulting in the absence of an interruption lag. Without this time span, there was no opportunity to explicitly create environmental cues or apply rehearsal before turning to the interrupting task. In consequence, naturally existing cues from memory or environment had to be used for resumption, e. g. the memorized selection content or visible selection marks. The modeled task always started with reading the written product list and remembering its content. Products were then selected within a navigation and selection procedure spanning the three menus illustrated in Figure 2.

In case an interruption occurred, its message was read and a random decision for or against the offered product was made. By performing either a memory retrieval of the previously selected product or looking for a relevant environmental cue, i. e. a selection mark linked to the lastly selected product, resumption started. Next a retrieval of the selection history, i. e. products already selected within the run and previously stored in the intermediate memory, was attempted. This is reminiscent of the memory-for-problem-states theory recently reported by Borst, Taatgen and van Rijn (2015). Afterwards, the next product was then searched for or the run ended when the reconstruction of the previous selection failed. A run ended through either successfully selecting all four products or failing to retrieve the next product. After completing the third and last run, the final product recall occurred. Under conditions of enhanced memory demands, the product-related person had to be remembered throughout the task and recalled at the end as well.

2.4 Experimental Validation

To assess the adequacy of the cognitive user model in terms of human behavior, an experimental validation was conducted. 62 participants aged 20 to 49 years (M = 28.53, SD = 7.16 years) performed the described task using a LG Google Nexus 4 smartphone with a screen size of 4.7", a display resolution of 1280 x 768 pxl, a pixel density of 319 ppi and Android 4.4.2 (KitKat) serving as operating system.

A shopping scenario was created to foster the participant’s involvement in the task. It asked them to imagine being a virtual character that does shopping with aid of the app during a usual day in life. Under conditions of enhanced memory demands two additional characters were introduced, an old neighbor and a sick friend. Participants were provided with information on attributes and habits of both people as well as their relationship to the main character. Since the latter was distinguished as caring and social person, shopping assistance for needy fellows was regarded as natural behavior and suitable for enhancing the external validity of the task.

3.1 Model Behavior

Looking at the resulting model behavior, when there were interruptions, fewer recalled products were shown and product selection took slightly longer. As can be seen in Table 1, such effects showed up especially with increasing frequency of interruption. Moreover, without enhanced memory demands, the model performed better across all performance-related parameters.

3.2 Model Comparison

Based on the inspected model behavior, a decline in task performance with increasing interruption and mental demands was expected in human data as well. Descriptive values are displayed in Table 2, statistical effects concerning differences in terms of interruption and mental demands were inspected by computing either analyses of variance (ANOVAs) in the case of product selection time, selected products and resumption time or a χ² -test in the case of final recall performance. In summary, none of the ANOVAs achieved significant results, neither for interruption nor mental demands. In contrast, a significant difference between high and low mental demands in the case of final recall showed up, χ² (15, N = 62) = 25.397, p = .045, supporting the assumption of worse final recall performance when mental demands are enhanced.

According to Schunn and Wallach (2005), for purposes of model comparison, human and model data points for high respective low mental demands, as displayed in Table 1 and Table 2, were compared by visual as well as numerical means.

The graphic comparison is shown in Figure 3 and Figure 4. Obviously, apart from the amount of selected products, human data remain on a continuously higher level for all displayed parameters, indicating the model predominantly performed better than the human sample. For product selection time, in the case of enhanced mental demands, model and human data point towards a similar direction, marginally increasing with enhanced frequency of interruption. In conditions with low mental demands, model data form a nearly straight line but human data show a considerable difference with increasing interruption frequency. Regarding the number of selected products, human data stay almost at the same level across interruption frequencies for both kinds of mental demands but model data perform differently. Whereas there is a noticeable decrease with increasing interruption frequency for high mental demands, a decrease towards highly interrupted trials occurs for low mental demands. Considering resumption time, the visual comparison indicates a high similarity between model and human data for both levels of mental demand, although deviation between both datasets is slightly higher under conditions with low mental demands.

Figure 3

Comparison of model and human data concerning time, products and resumption. Error bars represent 95 % confidence intervals on human data.

Figure 4

Comparison of model and human data concerning final recall. Error bars represent 95 % confidence intervals on human data.

In terms of final recall performance, the model performs considerably better in both conditions but model and human data exhibit an equivalent trend with a significantly higher amount of recalled products under low mental demands.

For numerical comparisons, the commonly applied root mean squared deviation (RMSD) was used. It assesses the mean deviation between model and human data points regarding their exact location in units of the respective scale. Taking into account the scale’s extremities, smaller values indicate less distance between both datasets, i. e. point to a better fit. Results of the comparison, separated by the level of mental demands, are shown in Table 3. As displayed, e. g. in terms of resumption time under high mental demands, model and human data are located quite close together, whereas in terms of final recall, deviation is quite substantial.