Determination of the Proper Rest Time for a Cyclic Mental Task Using ACT-R Architecture

Abstract

Objective

Analysis of the effect of mental fatigue on a cognitive task and determination of the right start time for rest breaks in work environments.

Background

Mental fatigue has been recognized as one of the most important factors influencing individual performance. Subjective and physiological measures are popular methods for analyzing fatigue, but they are restricted to physical experiments. Computational cognitive models are useful for predicting operator performance and can be used for analyzing fatigue in the design phase, particularly in industrial operations and inspections where cognitive tasks are frequent and the effects of mental fatigue are crucial.

Method

A cyclic mental task is modeled by the ACT-R architecture, and the effect of mental fatigue on response time and error rate is studied. The task includes visual inspections in a production line or control workstation where an operator has to check products’ conformity to specifications. Initially, simulated and experimental results are compared using correlation coefficients and paired t test statistics. After validation of the model, the effects are studied by human and simulated results, which are obtained by running 50-minute tests.

Results

It is revealed that during the last 20 minutes of the tests, the response time increased by 20%, and during the last 12.5 minutes, the error rate increased by 7% on average.

Conclusion

The proper start time for the rest period can be identified by setting a limit on the error rate or response time.

Application

The proposed model can be applied early in production planning to decrease the negative effects of mental fatigue by predicting the operator performance. It can also be used for determining the rest breaks in the design phase without an operator in the loop.

Keywords

rest breaks mental fatigue cognitive task ACT-R architecture

Introduction

Mental fatigue is defined as a subjective feeling of tiredness or lack of energy after prolonged exposures to cognitively demanding activities (Boksem & Tops, 2008; Marcora, Staiano, & Manning, 2009). It can impair human performance and is thus one of the major factors that influence output in terms of quality and quantity (Meijman, 1997; Mital, Bishu, & Manjunath, 1991). Human performance is defined as the accomplishment of a task in accordance with agreed on standards of speed, accuracy, and attentional demand (Wickens, Hollands, Banbury, & Parasuraman, 2015).

It is generally accepted that mental fatigue is related to human attention (Gonzalez, Best, Healy, Kole, & Bourne, 2011; Jongman, 1998; Steinborn, Flehmig, Westhoff, & Langner, 2009). Fatigued workers have difficulty in focusing their attention, ignoring irrelevant information, and providing appropriate answers (Boksem, Meijman, & Lorist, 2005; Yang, Xiao, Liu, Wu, & Miao, 2013). To avoid prolonged exposures to cognitively demanding activities and eliminate the negative effects of fatigue, for example, fall risks, on the quality and quantity of cognitive output, frequent rest breaks are recommended (Lew & Qu, 2014).

It can be said that mental fatigue is caused by two main factors. Long periods of wakefulness and circadian rhythmicity influence the overall level of individual cognitive functioning (Gunzelmann & Gluck, 2008; Lisper & Kjellberg, 1972). The other main reasons for operator fatigue are information overload, concentration, monotony, and repetition (Barker & Nussbaum, 2011).

Mental fatigue can be analyzed via physiological means such as electroencephalograms (EEGs) (Jagannath & Balasubramanian, 2014; Okogbaa, Shell, & Filipusic, 1994; Shen, Ong, Li, Hui, & Wilder-Smith, 2007; Zhao, Zhao, Liu, & Zheng, 2012) or event-related potentials (ERPs) (Boksem et al., 2005; Murata, Uetake, & Takasawa, 2005; Yang et al., 2013). Other physiological data are also used to measure mental workload, for instance, galvanic skin response, heart rate, and respiration changes (Kim, Bang, & Kim, 2004; Picard, Vyzas, & Healey, 2001; Wagner, Kim, & André, 2005). Subjective measures such as self-report questionnaires (Meshkati & Hancock, 1988) or fatigue impact scales (Borg, 1982; Fisk et al., 1994) can also be applied. These methods are useful when applied to existing physical samples of a system. However, it would be difficult to predict the effect of fatigue during the early stage of system development by using them. Biomathematical models are also used to evaluate mental fatigue, especially that caused by sleep loss (Folkard & Tucker, 2003; Klerman & Hilaire, 2007; Mallis, Mejdal, Nguyen, & Dinges, 2004). They can predict overall level of functioning but do not generate predictions of performance in particular tasks (Gunzelmann & Gluck, 2008).

Computational cognitive models can help to analyze the effects of fatigue and also serve as prediction tools for performance level (Gonzalez et al., 2011). In this paper, the Adaptive Control of Thought Rational (ACT-R) (Anderson, 1993) cognitive architecture has been used to define human cognitive processes quantitatively. It can predict response time as well as human error to analyze the effects of mental fatigue. The learning mechanism in ACT-R together with the production system, which lets only one production rule to be executed at a time, and the various outputs are three valuable advantages that make this architecture more suitable among other cognitive architectures for the purpose of this research. In addition, these properties help to make the model closer to reality for simulation of cognitive behaviors.

Initial research in this area was done by Jongman (1998). She proposed an ACT-R model of mental fatigue in the Sternberg memory search task (Sternberg, 1969). In her paper, the effect of fatigue was evaluated by different and discrete values of the W and G parameters. W was defined as the inhibition of interfering processes and stimuli while G was defined as motivation. More description of these parameters will be given in the following sections. The task that was investigated by Jongman was neither repetitive, prolonged, nor monotonous. Therefore, fatigue effect could not be examined thoroughly.

Another study of computational models for fatigue has been reported by Gunzelmann and colleagues (Gunzelmann, Byrne, Gluck, & Moore, 2009; Gunzelmann & Gluck, 2009; Gunzelmann, Veksler, Walsh, & Gluck, 2015). They tried to model fatigue effect as a result of sleep deprivation rather than work duration or workload. Later on, they developed an ACT-R model focusing on signal duration and vigilance decrement caused by sleep loss (Gartenberg, Veksler, Gunzelmann, & Trafton, 2014).

One of the most recent studies to consider mental fatigue in the computational model was undertaken by Gonzalez and colleagues (2011). They developed a cognitive model for a data entry task and examined learning and fatigue effects simultaneously. In this model, they demonstrated the effects of arousal parameter G and attention control W on the response time and accuracy of the task. The decreased attention control W and arousal G, combined with a production compilation mechanism, produce simultaneous speedup and increase in errors. Thus, they inferred that fatigue does not have any negative effect on response time but increases the error rate during data entry tasks. Their finding regarding response time cannot be reliable because increased response time after a repetitive and prolonged task has been accepted by other studies (Bertelson & Joffe, 1963; Bunce, Warr, & Cochrane, 1993; Jongman, 1998; Kato, Endo, & Kizuka, 2009; Langner, Steinborn, Chatterjee, Sturm, & Willmes, 2010; Sanders & Hoogenboom, 1970; Steinborn et al., 2009).

Following previous studies, we propose a new ACT-R model of mental fatigue that represents a typical unvarying, prolonged, and cyclic task of visual inspection in industrial production lines. In our model, the W parameter, which is defined for attention control, is systematically decreased per cycle, and a production compilation mechanism is enabled to model the learning effect. This effect is present all along the tests; however, in this study, fatigue effect on response time and error rate are focused and discussed.

In what follows, we will detail our account of the impact of prolonged and repetitive work on attention within the ACT-R architecture. We will describe our cognitive model in ACT-R and apply the fatigue effects by the W parameter modulation. To validate the model, simulation and experimental results are collected and compared. Then the effect of fatigue on response time and error rate is analyzed by running the tests. The model can be used to determine the right start time for the rest pause to reduce the negative effects of fatigue. Finally, a sensitivity analysis of the degradation rate is performed, and the robustness of the model is discussed.

ACT-R Architecture

ACT-R is a theory for simulating and understanding human cognition that can computationally predict the response time or errors of operations (Jo, Myung, & Yoon, 2012). ACT-R V6.0 consists of eight modules, including (1) the goal module that holds the goal of the behavior; (2) the declarative module that holds human knowledge; (3) the procedural module, which is represented in IF-Then form and determines the events that should occur in the specific conditions; (4) the imaginal module that keeps information temporarily; (5) the visual and aural modules that process and encode the visual and auditory information from the external environment; and (6) the vocal and manual modules, which do actions in external environment when they receive commands from the procedural module (Anderson, 2007; Oh, Jo, & Myung, 2014).

On the other hand, ACT-R combines a symbolic level with a subsymbolic system. The symbolic level is an abstract characterization of how brain structure encodes knowledge. The subsymbolic level involves a set of mathematical procedures that determine access to the symbolic structure (Anderson, 2007).

Based on the studies mentioned, the effect of fatigue and learning in prolonged and repetitive cognitive task is attributed to two ACT-R subsymbolic factors. Lack of attention control is shown by W, and skill acquisition in combination with the production compilation mechanism is shown by production utility, which are explained in detail in the next sections.

Activation

In ACT-R mechanisms, interactions between modules take place in the form of chunks, which are structured units bundling a small amount of information. Each chunk in the declarative module has an activation value that is obtained on the basis of Equation 1:

A_{i} = B_{i} + \sum_{j} w_{j} S_{j i} + ε .

Activation $A_{i}$ of a chunk i reflects an estimate of how likely it is that the chunk would match a production at the current time as well as its speed of retrieval (Anderson & Lebiere, 1998; Anderson, Reder, & Lebiere, 1996). Activation of a chunk is determined by the base-level activation B_i, the associative activation $S_{j i}$ , and noise ε. However, there is one constraint on that. There is a parameter τ called the retrieval threshold, which is determined as an input to the model. The retrieval process of chunk i is possible when $A_{i} \geq τ$ .

The base-level activation B_i of the chunk i reflects the recency and frequency of using the chunk, which are calculated by Equation 2:

B_{i} = \ln (\sum_{j = 1}^{n} t_{j}^{- d}),

where n is the number of presentations for chunk i, $t_{j}$ is the time since the jth presentation of chunk i, and d is the decay rate, which is considered as an input data to the model.

This equation explains how the activation varies according to the usage history of the chunk. $S_{j i}$ reflects the impact of contextual values on chunk activation. This parameter between two chunks j and i is 0 if chunk j is not related to chunk i and chunk j and i are not the same. The parameter ε is a variable noise value associated with a chunk and can be interpreted as environmental changes that affect the retrieval likelihood.

$w_{j}$ represents the attention weight of each of the elements j that are part of the current goal chunk. ACT-R assumes that W, or the overall attention that one can distribute over source objects and achieve via the sum of attentions $w_{j}$ , is limited (Anderson et al., 1996, 2004). W is a parameter that is thought to reflect individual differences in the ability to make memory retrieval context-sensitive. This parameter helps to create contrast between relevant and irrelevant information for the current goal (Reder, 1997). Lower values of W reduce activation, increasing the possibility that incorrect items will be retrieved from memory (Gonzalez et al., 2011). Activation of a chunk also determines how quickly it can be retrieved. When a retrieval request is made, the time it takes until the chunk retrieved is achieved by Equation 3:

T = F e^{- A},

where T is the time, A is the activation of the retrieved chunk, and F is the latency factor, which is set as input to the model. Based on this equation, it can be understood that the higher the level of activation, the faster the retrieval.

Utility Mechanism and Learning

In ACT-R, the strength-like quantity that determines what rules fire is often referred to as utility because it is a measure of the value of the rule (Anderson, 2007). When a model finds itself in a situation where multiple rules can be applied, it chooses a rule with the highest utility.

In ACT-R 6.0, a new utility mechanism has been included, called temporal difference (TD algorithm). It is fundamentally different from the utility mechanism in ACT-R 5.0. The TD algorithm is based on reinforcement learning by which production compilation generates new production rules based on existing ones (Gonzalez et al., 2011). Separate learning mechanisms are used in the architecture, but the so-called G parameter is not used in learning mechanism.

The proposed model is not pretrained, but the learning mechanism is enabled and the retrieval process is improved while repeating the task. In the long term, when the model reaches its steady state, the learning effect would be negligible. In contrast, the fatigue effects will be more significant and could affect performance. Therefore, the current model does not investigate the effect of learning independently but focuses on fatigue effects during the prolonged and repetitive simulated task.

Fatigue Effect in ACT-R

A common method of obtaining fatigue effects, which is also used in the present study, involves decrements in W, which is associated with attention in ACT-R. In the ACT-R mechanism, decreased activation produces more errors and directly relates to reduction of the speed of fact retrieval (Anderson et al., 2004).

Other alternative mechanisms in ACT-R that may potentially capture the effects of prolonged work are presented by Fu and Gonzalez (Fu, Gonzalez, Healy, Kole, & Bourne, 2006; Gonzalez, Fu, Healy, Kole, & Bourne, 2006) and Gunzelmann (Gunzelmann, Gluck, Moore, & Dinges, 2012). In the Fu and Gonzalez studies, they increased activation noise within ACT-R combined with the motivational parameter, G. Even though it is reasonable to change the noise, the W parameter has a clearer correspondence to cognitive variables involved in prolonged work compared to activation noise. The direct correspondence of noise to cognitive constructs is less clear and cannot be interpreted well as a human cognitive tool (Gonzalez et al., 2011). In Gunzelmann’s research, the effect of sleep deprivation has been modeled by a scaling parameter F_d or fatigue-declarative as a multiplicative moderator of base-level activation, which ranges from 0 to 1 to decrement the activation. So they modified Equation 1 to Equation 4.

A_{i} = F_{d} (B_{i}) + \sum_{j} w_{j} S_{j i} - D_{i p} + ε .

In their model, the declarative knowledge is mostly well practiced and highly available. Therefore, the base level activation does not change considerably during the experiments, and it is logical to use a detrimental factor for based level activation. However, in the current research, the base level activation is dynamic itself and increases while repeating the tasks. Therefore, W degradation is more reasonable to consider mental fatigue for the proposed task.

In order to obtain fatigue effects in this research, the W parameter is defined to decrease gradually from its initial value. A reduction multiplier with a fixed value is defined, which comes into effect in each cycle. The initial value of W is assumed to be 0.95, and the W degradation occurs at the rate of .00125 per second. Other parameters such as retrieval threshold (τ), latency factor (F), and the instantaneous noise distribution parameter (s) are set to 0.01, 0.8, and 0.5, respectively. The basic method to determine model parameters are based on previous studies (Anderson, 1974; Zbrodoff, 1995). The decay parameter gets its default value. Instantaneous noise distribution parameter and initial value of W are taken from background studies (Anderson, 1981; Gonzalez et al., 2011). In addition, a sensitivity analysis has been performed to determine the final settings of these parameters.

Retrieval threshold (:rt), latency factor (:lf), and W degradation are calculated through independent iterative processes. In each iteration, one of the parameters is assigned a value taken from ACT-R models in the repository while other parameters are remained constant. In each iteration, model output (response time and error rate) are analyzed for different task cycle time, number of cycles, and product types. Then root mean square deviation (RMSD) and r coefficient are computed so simulated and human results can be compared. At the end of each iteration, RMSD values for response time and error rate are added together, and r values for response time and error rate are added together as well. Among the four maximum sum of r values, the parameter levels, which correspond to minimum sum of RMSD, are selected as the iteration solution. This solution forms the initial setting for the next iteration. The whole process has been run for four times and then stopped because of the proximity of the third and fourth iteration solutions.

Method

The proposed model represents typical control activities in production lines in industry or mail services. In some production lines, products pass in rapid succession through control stations where the operators visually check the tags’ conformity to predefined specifications.

For the control task in this research, it is assumed that products are classified in four different types. For further comparison, a three-type model is also discussed in the following.

Four colors—red, blue, dark-grey, and yellow—are assigned to product types 1, 2, 3, and 4, respectively, corresponding to tags 1, 2, 3, and 4. The operator has to check whether the color of each product matches its tag. Figure 1 shows this configuration, which is also restored in the form of chunks in the declarative memory of the ACT-R model.

Figure 1.

Input data for the model.

Depending on the speed of the production line, the control operation of each product must be completed within a task cycle time. The term task cycle time should not be confused with duration of the matching-selection-execution cycle in ACT-R’s procedural knowledge. During each task cycle time of the test, a new product with a random color and label from the available range is presented to the operator.

When the correct tag is retrieved from memory and compared with the present tag, the operator must take the appropriate action according to his or her accept or reject decision. In the computer model, if the tags are identical, the G key on the keyboard is pressed; otherwise the H key is pressed. As this is the only physical part of the activity, the whole task can be well presented by a cognitive ACT-R model. This is called a control model and can be run either by a human participant or automatically by computer simulation. The control model is shown by a flowchart in Figure 2.

Figure 2.

Flowchart of the control model.

Participants

Two groups of volunteers participated in the tests. The first group (Group 1) was used to verify the proposed model and consisted of 10 healthy young adults (5 males and 5 females). All participants were 21 to 26 years old to minimize potential age-related differences. The mean (SD) ages of the male and female participants were 25.6 (0.54) and 25.4 (1.1) years, respectively.

The second group (Group 2) was used to analyze fatigue in the simulated cognitive task. A total of 20 healthy young adults, 8 males and 12 females, with similar educational and work background, performed the simulated task for 50 minutes. The mean (SD) ages of participants were 25.25 (2.9) and 23.53 (1.8) for males and females, respectively. The participants were all university students because quality control operators were not available to do the tests. Environmental conditions and timing of the tests were also kept consistent for all participants. Nevertheless, individual differences cannot be completely cancelled out.

In the experimental setup of the simulated task, a P4 personal computer with a 19-inch LCD monitor and a standard keyboard was used. Each time the task was restarted, a box with a random color and label from the available range was displayed on the screen (Figure 3). The participant had to decide quickly and press G or H on the keyboard before the task cycle time was over. If the operator pressed the key before the task cycle time elapsed, the time was saved as the response time. Otherwise, the test failed, an omission error occurred, and the task cycle time was considered as response time. If the operator pressed the wrong key, an error of commission was recorded. This type of error does not happen in simulation tests. By the end of each task cycle time, the next box will be displayed without any break. In what follows, the data obtained from the participants are called human results.

Figure 3.

A sample test display.

ACT-R Model

Among the eight modules of ACT-R mentioned previously, four modules, including declarative, imaginal, visual, and manual modules, together with the production system for integrating and processing modules were used in the control model. The ACT-R v6.0 model was programmed in Common Lisp version 1.8-r15286M, and its main production rules are shown in Table 1.

Table 1:

Production Rules of the Control Model

Production Rule No.	Title	If	Then
1	Identify-type-of-product	Goal is “find-type-of-product,” and the product with color X is in the visual module	Retrieve type of product with color X
2	Identify-correct-tag	Goal is “find-appropriate-tag,” and the product with color X and type Y is in the retrieval module	Retrieve the correct tag based on the type Y
3	Checking-conformity	Goal is “checking-conformity,” and the correct tag and piece tag are in the imaginal module	Check whether the product’s tag is equal to the correct tag
4	Response-conforming	Goal is “response,” and the product’s tag is equal to the correct tag	Press the G key and change goal to “start”
5	Response-nonconforming	Goal is “response,” and the product’s tag is not equal to the correct tag	Press the H key and change goal to “start”

An example of retrieving the correct tag based on the quality level is depicted in Figure 4. The spread activation starts from the goal module, and the chunk that is stored in the buffer of the goal is a Usage type with one slot that is filled with High quality. Therefore $W = w_{1} when n = 1$ and the model searches chunks with Usage type, and the activation for these chunks is calculated. As illustrated in Figure 4, base level activation and noise parts are the only parts considered for those chunks that do not have High quality in their slots. In other words, $S_{1 i} = 0 when i \neq 1$ for all other existing chunks in the declarative module. As the time passes and the model gets fatigued, $w_{1}$ is reduced, and consequently the activation of the selected chunk is decreased.

Figure 4.

The process of calculating activations.

Error occurs when no chunk has an activation value above the retrieval threshold or it is so low that it takes more than the specified task cycle time. Therefore, mental fatigue may affect production rules No. 1 and 2, reducing the activation of the correct chunk. Therefore, it increases errors and response time based on Equations 1 and 3.

Results and Discussion

Verification of the Model

By running the computer program in eight different settings, the simulated and human (Group 1) results were generated. Each setting was identified by its task cycle time (1.5 seconds or 3 seconds), number of cycles (250 or 500), and type of product (3 or 4). The participants of Group 1 conducted the eight test settings in eight different days at the same time of the day. Based on the power calculations for stable predictions (Howell, 2012; Ritter, Schoelles, Quigley, & Klein, 2011), for the power greater than 0.7, the number of model runs was set to n = 30. The results are summarized in Table 2 and depicted for response time in Figure 5 and error rate in Figure 6. It should be noted that the average response time was obtained by dividing the total response time by the number of cycles. The average commission error for human results is different for each test setting, but it does not exceed .092 and in most cases is less than omission error. In addition, it must be mentioned that the model does not include physical fatigue or other stressors such as task importance.

Table 2:

Results of the Control Model (Times Are in Seconds)

				Mean (Standard Deviation) Response Time		Mean (Standard Deviation)Error Rate
Test Setting	Task Cycle Time	Number of Cycles	Types of Products	Human	Simulation	Human	Simulation
1	3	250	3	1.414 (0.26)	1.353 (0.163)	0.129 (0.014)	0.122 (0.087)
2	3	250	4	1.488 (0.21)	1.471 (0.107)	0.138 (0.022)	0.157 (0.058)
3	3	500	3	1.353 (0.102)	1.239 (0.044)	0.063 (0.037)	0.082 (0.081)
4	3	500	4	1.339 (0.014)	1.368 (0.012)	0.064 (0.026)	0.108 (0.026)
5	1.5	250	3	0.840 (0.034)	1.073 (0.025)	0.121 (0.011)	0.112 (0.059)
6	1.5	250	4	1.002 (0.103)	1.127 (0.038)	0.209 (0.024)	0.170 (0.082)
7	1.5	500	3	0.952 (0.21)	1.015 (0.090)	0.142 (0.078)	0.080 (0.032)
8	1.5	500	4	0.967 (0.12)	1.109 (0.03)	0.179 (0.115)	0.146 (0.061)

Figure 5.

Average task cycle’s response time.

Figure 6.

Average error rate.

As shown in Figure 5, when there are more types of products, in most cases, the average response time increases (Test Setting 1 vs. 2 or Setting 5 vs. 6). That is because the number of items the participant must remember has increased. In extreme cases, where many types of products are presented, information overload may occur. In the ACT-R view, when the number of chunks in the declarative memory increases, the number of presentations for each chunk is generally decreased so that the base level activation of chunks is reduced. As a result, the overall activation of chunk is decreased, and based on Equation 3, the response time for retrieval increases. This interpretation can also be true for the average error rate in Figure 6.

The main source of differences in response times both in human and simulated data is the task cycle time. Response times are clearly longer when task cycle times are longer. Because of time pressure, it is expected that the average error rate increases when task cycle time is decreased at the same number of cycles and types of products. This can be seen in Figure 6, especially when more product types are checked (Test Setting 2 vs. 6 or Setting 4 vs. 7).

In Figures 5 and 6, both average task cycle’s response time and average error rate at setting with 500 cycles are less than settings with 250 cycles (Test Setting 1 vs. 3 and 2 vs. 4). That happens mainly because of the learning effects. It is worth mentioning that for 500 cycle settings, the maximum time of the test does not exceed 25 minutes, hence the learning that occurs produces improvements in performance that exceed the fatigue effects early in the task.

Goodness-of-fit analysis was performed, and root mean square deviations, correlation coefficients, and p values were calculated; namely, RMSD_{response time} = 3.822E-06, r_{response time} = 0.93, p = 0 for response time; RMSD_{error rate} = .001, r_{error rate} = 0.72 and p = .043 for error rate. In addition, a paired t test for response time (t = −1.11, p = .303) showed no significant difference between the human and the simulation results (Montgomery, 2007). Moreover, all t statistic values lay within a 95% confidence interval of −2.36 and 2.36. This was also true for the error rate (t = 0.870, p = .413), with a 95% confidence interval of −2.36 and 2.36. Since nonsignificance may simply be due to large error variance or insufficient statistical power, multivariate analysis of variance of the results are also performed for response time, and error rate and showed no statistically significant difference in neither human nor model output, namely, F(2, 13) = 0.139, p > .05; Wilk’s Λ = 0.979. These findings lead to the conclusion that the control model can predict the performance in different situations and is a reliable means of evaluating mental fatigue related to this cognitive task.

Evaluating the Fatigue Effect

To evaluate the mental fatigue effect, tests were run for 1,000 cycles, with a task cycle time of 3 seconds and four different types of product. Each run of test lasted 50 minutes and were done both by human participants (Group 2) and computer simulation. Similar to the verification tests described previously, the number of model runs was set to n = 60 for the power greater than 0.9. The results are graphically depicted in Figures 7 and 8 for response time and error rate. Each point on Figure 7(a) represents the average of the response time at each cycle for 20 participants. In Figure 7(b), each point represents the average response time at each cycle for 60 individual test results. The same settings are used for error rate results in Figure 8. To show the separate effect of fatigue and learning, the simulated model was run once for each effect when another effect was off. The results are depicted in Figures 7(b.1) and 7(b.2) for response time and 8(b.1) and 8(b.2) for error rate.

Figure 7.

The impact of mental fatigue on response time. (a) Human test results, (b) simulated test results, (b.1) simulated test results without fatigue effect, and (b.2) simulated test results without learning effect.

Figure 8.

(a) Impact of mental fatigue on error rate. (b.1) Simulated test results without fatigue effect. (b.2) Simulated test results without learning effect.

As shown in Figures 7(a) and 7(b), fatigue in both simulated and human tests has a negative impact on response time and increases the value of this variable about 0.62 seconds (20% of task cycle time) in the last 300 cycles (from 30 to 50 minutes). Figure 8(a) also shows the negative impact of fatigue on error rate. Effects of fatigue in simulated and experimental tests are apparent, with a positive trend at about 750 cycles (37.5 minutes) and 700 cycles (37 minutes) for simulated and human results in Figure 8(a). On average, the fatigue effect causes the error rate to increase by 7%. This is in agreement with DeMarco and Lister’s (1988) research in which it was suggested that the majority of adults are unable to sustain attention beyond 40 minutes.

In this model, the learning mechanism mostly affects early stages of the test. This can be seen in Figures 7(b.2) and 8(b.2), which show the result of the tests when production compilation was off. The initial trends in these figures are lower compared to corresponding Figures 7(b) and 8(a), where the status of the production compilation was on. In addition, the fatigue has more effect as cycle number increases at the end of the test. If we eliminate the fatigue effect as can be seen in Figures 7(b.1) and 8(b.1), error rate will decrease and tend to zero, and response time will tend to a minimum value that is not reasonable.

By comparing Figures 7(a) and 7(b), it can be inferred that the human results are more scattered than the simulated results: The variance of the human results is .089, whereas that of the simulated results is .0480. This may be due to the individual differences reported previously. On the other hand, for correlation tests of the data presented in Figures 7(a) and 7(b), RMSD = .092, r = 0.545, and p < .001, which shows similar trends during the 50 minutes of the test. This can be interpolated by simple polynomial regressions of order 2 by Equations 5 and 6:

y = 2 E - 06 x^{2} - . 0015 x + 1.7573, for Figure 7 (a)

y = 2 E - 06 x^{2} - . 0017 x + 1.7497, for Figure 7 (b)

In these equations, y represents average response time, and x is the number of cycles. The same trend with RMSD = .001, r = 0.596, and p < .001 similarity is true for error rate in Figure 8(a). The average commission error for human results is equal to .042, which is less than average omission error. In both Figures 7 and 8, at a low number of cycles, response time and error rate gradually decline, but after a while, the trends are reversed, and both variables start to incline. This can be justified by the learning phase that exists both in simulated and experimental tests.

The difference between human and simulated error rate at each cycle is plotted in Figure 9. It never exceeds 0.134 and gradually decreases, which shows that the simulated model works acceptably.

Figure 9.

The diffences between human and simulated error rate results.

Determination of the Rest Time

The control model can be used in the planning stage to identify the required rest pauses during working hours for the particular task of this research. Regarding the acceptable error rate or agreed rate of nonconformities, it is possible to relate it to the mental pressure on the operator.

In the proposed model, if the task cycle time is set at 3 seconds and the acceptable error rate at 6%, it can be inferred from Figure 10 that the operator could work acceptably until about 804 cycles or 40 minutes. After that, he or she becomes overfatigued, and the error rate exceeds the acceptable limit. If the operator does not take a rest and continues to work, after 1,100 cycles or 55 minutes, the error rate will surpass 10%, which means that 4% of the products might be wasted. Similarly, the optimum task cycle time can be determined and implemented by adjusting the speed of the transfer production line. It should be noted that these conclusions cannot yet be generalized to other tasks unless validated individually.

Figure 10.

Rest time for the sample cognitive task.

Sensitivity to W Degradation Rate

Lack of planned rest breaks could cause mental mistakes, which become more frequent during prolonged time on task (Steinborn et al., 2009). As a result, participants cannot answer as swiftly as possible when fatigued. This is included in the model by defining W degradation per second in each cycle, which results in decreased activation of the related chunk. To analyze the sensitivity of the model to W degradation level, the control model is run 30 times for 1,000 cycles, with a task cycle time of 3 seconds and four different types of product.

For six different values of W degradation, the average response time and error rate are shown in Figures 11 and 12. These figures demonstrate that the model is mathematically stable around the predefined rate of .00125, as stated previously. At rates lower than .001 or higher than .005, the average response time and error rate are not very sensitive. At a high rate of W degradation, the W parameter decreases considerably, and hence the activation of related chunk reduces significantly. As a consequence, the likelihood of retrieval is decreased, and errors occur more frequently. In addition, decreased activation increases both the time required for retrieval and the response time.

Figure 11.

Response time sensitivity to W degradation.

Figure 12.

Error rate sensitivity to W degradation.

Conclusion

This research has evaluated the effects of mental fatigue on performance degradation to determine the right start time for the rest pause. An ACT-R model of a prolonged and cyclic cognitive task was built and verified by experimental tests. The model can represent visual inspection in a production line or control workstation where an operator has to check products’ conformity to specifications. Since the fatigue effect is not yet included in ACT-R architecture, a degradation rate was defined in the computer program of the model. By running this program both by human participants and simulation, the negative effect of the fatigue was apparent in 50 minutes of tests. It was revealed that during the last 20 minutes of the tests, the response time increased by 20%, and during the last 12.5 minutes of the tests, the error rate increased by 7% on average. A sensitivity analysis of the degradation rate showed that the model is nearly stable in the vicinity of the defined value, namely, .001 per second. Future studies can investigate the length of the rest breaks and recovery rate and evaluate the influence of gender, age, and environmental factors on the results.

Key Points

A cyclic mental task was modeled in ACT-R to investigate the response time and error rate.

Using the proposed model, the start time for rest breaks can be determined.

To be validated and tested, the model was run both by simulation and using human operators.

Human data and simulations show average increase of 20% in response time and 7% in error rate over the 50-minute work period.

Footnotes

Nooshin Atashfeshan received her BSc and MSc degrees in industrial engineering from the Faculty of Engineering, Ferdowsi University of Mashhad, Iran in 2011 and 2013, respectively. Her current research interests include mental fatigue, cognitive science, ACT-R cognitive architecture, rest allowances, and MIS systems. Her MSc thesis analyzed the effect of mental fatigue on human performance, especially in industrial operations. She is a member of the Institute for Cognitive Science Studies and has published a research paper on ignoring the fatigue effect in performance measurement.

Hamideh Razavi received her PhD in mechanical engineering from Imperial College London in 2006. Her undergraduate degrees were in industrial engineering from Sharif University and Tehran Polytechnic, Iran. She joined the Industrial Engineering Department of Ferdowsi University of Mashhad as an assistant professor in 2007 and became an associate professor in 2013. Following her experiences in the automotive industry, she has taught courses in industrial human factors and published several papers on both physical and cognitive ergonomics. Her current research interests are mental workload and fatigue measurement, vigilance assessment and rating, and aesthetics in product design.

References

Anderson

J. R.

(1974). Retrieval of propositional information from long-term memory. Cognitive Psychology, 5, 451–474.

Anderson

J. R.

(1981). Interference: The relationship between response latency and response accuracy. Journal of Experimental Psychology: Human Learning and Memory, 7(5), 326.

Anderson

J. R.

(1993). Rules of the mind. Hillsdale, NJ: Erlbaum.

Anderson

J. R.

(2007). How can the human mind occur in the physical universe? New York, NY: Oxford University Press.

Anderson

J. R.

Bothell

Byrne

M. D.

Douglass

Lebiere

Qin

(2004). An integrated theory of the mind. Psychological Review, 111(4), 1036–1060.

Anderson

J. R.

Lebiere

(1998). The atomic components of thought. Mahwah, NJ: Erlbaum.

Anderson

J. R.

Reder

L. M.

Lebiere

(1996). Working memory: Activation limitations on retrieval. Cognitive Psychology, 30(3), 221–256.

Barker

L. M.

Nussbaum

M. A.

(2011). The effects of fatigue on performance in simulated nursing work. Ergonomics, 54(9), 815–829.

Bertelson

Joffe

(1963). Blockings in prolonged serial responding. Ergonomics, 6(2), 109–116.

10.

Boksem

M. A.

Meijman

T. F.

Lorist

M. M.

(2005). Effects of mental fatigue on attention: An ERP study. Cognitive Brain Research, 25(1), 107–116.

11.

Boksem

M. A.

Tops

(2008). Mental fatigue: Costs and benefits. Brain Research Reviews, 59(1), 125–139.

12.

Borg

(1982). A category scale with ratio properties for intermodal and interindividual comparisons. In H.-G. Geissler & P. Petzold (Eds.), Psychophysical judgment and the process of perception (pp. 25–34). Berlin: Deutscher Verlag der Wissenschaften.

13.

Bunce

D. J.

Warr

P. B.

Cochrane

(1993). Blocks in choice responding as a function of age and physical fitness. Psychology and Aging, 8(1), 26–33.

14.

DeMarco

Lister

1988. Peopleware: Productive projects and teams. Reading, MA: Addison-Wesley.

15.

Fisk

J. D.

Ritvo

P. G.

Ross

Haase

D. A.

Marrie

T. J.

Schlech

W. F.

(1994). Measuring the functional impact of fatigue: Initial validation of the fatigue impact scale. Clinical Infectious Diseases, 18(Supplement 1), S79–S83.

16.

Folkard

Tucker

(2003). Shift work, safety and productivity. Occupational Medicine, 53(2), 95–101.

17.

W. T.

Gonzalez

Healy

A. F.

Kole

J. A.

Bourne

L. E.

(2006). Building predictive human performance models of skill acquisition in a data entry task. Proceedings of the Human Factors and Ergonomics Society 50th Annual Meeting (pp. 1122–1126). Santa Monica, CA: Human Factors and Ergonomics Society.

18.

Gartenberg

Veksler

Gunzelmann

Trafton

(2014). An ACT-R process model of the signal duration phenomenon of vigilance. Proceedings of the Human Factors and Ergonomics Society 58th Annual Meeting (pp. 909–913). Santa Monica, CA: Human Factors and Ergonomics Society.

19.

Gonzalez

Best

Healy

A. F.

Kole

J. A.

Bourne

L. E.

(2011). A cognitive modeling account of simultaneous learning and fatigue effects. Cognitive Systems Research, 12(1), 19–32.

20.

Gonzalez

Healy

A. F.

Kole

J. A.

Bourne

L. E.

Jr. (2006). ACT-R models of training data entry skills. In Proceedings of the Fifteenth Conference on Behavior Representation in Modeling and Simulation (pp. 101–109). Orland, FL: SISO.

21.

Gunzelmann

Byrne

M. D.

Gluck

K. A.

Moore

L. R.

(2009). Using computational cognitive modeling to predict dual-task performance with sleep deprivation. Human Factors: The Journal of the Human Factors and Ergonomics Society, 51, 251–260.

22.

Gunzelmann

Gluck

K. A.

(2008). Approaches to modeling the effects of fatigue on cognitive performance. In Hansberger

(Ed.), Proceedings of the Seventeenth Conference on Behavior Representation in Modeling and Simulation (pp. 136–145). Orlando, FL: Simulation Interoperability Standards Organization.

23.

Gunzelmann

Gluck

(2009). An integrative approach to understanding and predicting the consequences of fatigue on cognitive performance. Cognitive Technology, 14(1), 14–25.

24.

Gunzelmann

Gluck

Moore

L. R.

Dinges

D. F.

(2012). Impaired performance due to sleep deprivation: A role for diminished knowledge access. Cognitive Systems Research, 13(1), 1–11.

25.

Gunzelmann

Veksler

Walsh

Gluck

(2015). Understanding and predicting the cognitive effects of sleep loss through simulation. Translational Issues in Psychological Science, 1(1), 106–115.

26.

Howell

(2012). Statistical methods for psychology. Boston, MA: Cengage Learning.

27.

Jagannath

Balasubramanian

(2014). Assessment of early onset of driver fatigue using multimodal fatigue measures in a static simulator. Applied Ergonomics, 45(4), 1140–1147.

28.

Myung

Yoon

(2012). Quantitative prediction of mental workload with the ACT-R cognitive architecture. International Journal of Industrial Ergonomics, 42(4), 359–370.

29.

Jongman

(1998). How to fatigue ACT-R. In Proceedings of The Second European Conference on Cognitive Modelling (pp. 52–57). Nottingham: Nottingham University Press.

30.

Kato

Endo

Kizuka

(2009). Mental fatigue and impaired response processes: Event-related brain potentials in a go/nogo task. International Journal of Psychophysiology, 72(2), 204–211.

31.

Kim

K. H.

Bang

S. W.

Kim

S. R.

(2004). Emotion recognition system using short-term monitoring of physiological signals. Medical and Biological Engineering and Computing, 42(3), 419–427.

32.

Klerman

E. B.

Hilaire

M. S.

(2007). Review: On mathematical modeling of circadian rhythms, performance, and alertness. Journal of Biological Rhythms, 22(2), 91–102.

33.

Langner

Steinborn

M. B.

Chatterjee

Sturm

Willmes

(2010). Mental fatigue and temporal preparation in simple reaction-time performance. Acta Psychologica, 133(1), 64–72.

34.

Lew

F. L.

(2014). Effects of mental fatigue on biomechanics of slips. Ergonomics, 57(12), 1927–1932.

35.

Lisper

H. O.

Kjellberg

(1972). Effects of 24-hour sleep deprivation on rate of decrement in a 10-minute auditory reaction time task. Journal of Experimental Psychology, 96(2), 287–290.

36.

Mallis

M. M.

Mejdal

Nguyen

T. T.

Dinges

D. F.

(2004). Summary of the key features of seven biomathematical models of human fatigue and performance. Aviation, Space, and Environmental Medicine, 75(Supplement 1), A4–A14.

37.

Marcora

S. M.

Staiano

Manning

(2009). Mental fatigue impairs physical performance in humans. Journal of Applied Physiology, 106(3), 857–864.

38.

Meijman

T. F.

(1997). Mental fatigue and the efficiency of information processing in relation to work times. International Journal of Industrial Ergonomics, 20(1), 31–38.

39.

Meshkati

Hancock

(1988). Human mental workload. In Hancock

P. A.

Meshkati

(Eds.), Advances in psychology (pp. v–vi). North-Holland: Elsevier Science Publishers B.V.

40.

Mital

Bishu

R. R.

Manjunath

S. G.

(1991). Review and evaluation of techniques for determining fatigue allowances. International Journal of Industrial Ergonomics, 8(2), 165–178.

41.

Montgomery

D. C.

(2007). Introduction to statistical quality control. New York, NY: John Wiley & Sons.

42.

Murata

Uetake

Takasawa

(2005). Evaluation of mental fatigue using feature parameter extracted from event-related potential. International Journal of Industrial Ergonomics, 35(8), 761–770.

43.

Myung

(2014). Computational modeling of human performance in multiple monitor environments with ACT-R cognitive architecture. International Journal of Industrial Ergonomics, 44(6), 857–865.

44.

Okogbaa

O. G.

Shell

R. L.

Filipusic

(1994). On the investigation of the neurophysiological correlates of knowledge worker mental fatigue using the EEG signal. Applied Ergonomics, 25(6), 355–365.

45.

Picard

R. W.

Vyzas

Healey

(2001). Toward machine emotional intelligence: Analysis of affective physiological state. Pattern Analysis and Machine Intelligence, IEEE Transactions, 23(10), 1175–1191.

46.

Reder

L. M.

(1997). Modeling individual differences in a digit working memory task. In Proceedings of the Nineteenth Annual Conference of the Cognitive Science Society (Vol. 19, p. 460). Mahwah, NJ: Lawrence Erlbaum Associates.

47.

Ritter

F. E.

Schoelles

M. J.

Quigley

K. S.

Klein

L. C.

(2011). Determining the number of simulation runs: Treating simulations as theories by not sampling their behavior. In Human-in-the-loop simulations (pp. 97–116). London: Springer.

48.

Sanders

A. F.

Hoogenboom

(1970). On the effects of continuous active work on performance. Acta Psychologica, 33, 414–431.

49.

Shen

K. Q.

Ong

C. J.

X. P.

Hui

Wilder-Smith

(2007). A feature selection method for multilevel mental fatigue EEG classification. Biomedical Engineering, IEEE Transactions, 54(7), 1231–1237.

50.

Steinborn

M. B.

Flehmig

H. C.

Westhoff

Langner

(2009). Differential effects of prolonged work on performance measures in self-paced speed tests. Advances in Cognitive Psychology, 5, 105–113.

51.

Sternberg

(1969). Memory-scanning: Mental processes revealed by reaction-time experiments. American Scientist, 57, 421–457.

52.

Yang

Xiao

Liu

Miao

(2013). Mental fatigue impairs pre-attentive processing: A MMN study. Neuroscience Letters, 532, 12–16.

53.

Wagner

Kim

André

(2005, July). From physiological signals to emotions: Implementing and comparing selected methods for feature extraction and classification. Paper presented at Multimedia and Expo, ICME.

54.

Wickens

C. D.

Hollands

J. G.

Banbury

Parasuraman

(2015). Engineering psychology & human performance. New York, NY: Psychology Press.

55.

Zbrodoff

N. J.

(1995). Why is 9 + 7 harder than 2 + 3? Strength and interference as explanations of the problem-size effect. Memory & Cognition, 23(6), 689–700.

56.

Zhao

Liu

Zheng

(2012). Electroencephalogram and electrocardiograph assessment of mental fatigue in a driving simulator. Accident Analysis & Prevention, 45, 83–90.