What Do Traffic Simulations Have to Provide for Virtual Road Safety Assessment? Human Error Modeling in Traffic Simulations

Will Advanced Driving Assistance Systems (ADAS) and Highly Automated Driving (HAD) perform in the expected manner? Will they actually make road traffic safer, or will they potentially introduce new critical situations or road accidents? It is almost impossible to address these questions solely through real-world tests. A promising tool to provide appropriate answers in a time- and cost-efficient way without exposing subjects to risk are virtual assessment methods. Reliable safety assessments are only possible, if the traffic simulations provide realistic traffic, including critical situations and road accidents. This paper provides a review of how human error contributes to critical situations and accidents in road traffic. The focus is on the causes and mechanisms of human error, which driver behavior models must address in order to simulate realistic traffic. For this purpose, Rasmussen’s error taxonomies are applied to the traffic context and extended with further research. The paper shows the causes of those human errors and that the underlying mechanisms thereof should be taken into account in order to obtain more transparent and realistic driver behavior models. It is shown, which concepts for modelling realistic traffic exist and how virtual safety assessment could benefit from this development. In addition, the driver behavior model DReaM (Driver Reaction Model) is presented to address the issues resulting from existing cognitive driver models.


I. INTRODUCTION
T ECHNOLOGIES, such as Advanced Driving Assistance Systems (ADAS) built into vehicles, have become increasingly complex in recent years, and this trend is continuing with regard to Highly Automated Driving (HAD). As a result, ADAS and HAD will cover more demanding application areas, and interactions with other road users will increase in the future (see [1], [2]), especially when ADAS and HAD are going to tackle urban traffic. Externally displayed lights or symbols could inform surrounding traffic of future movements or signal pedestrians and cyclists that they have been detected [3].
Manufacturers must ensure that their newly developed systems are safe and behave in the expected manner. This includes investigations into whether new systems cause new critical situations or even lead to potential accidents [4]. The validation of ADAS and HAD that address more extensive application areas, such as urban junctions or take over driving tasks, solely by real-world tests is becoming increasingly difficult.
The estimated number of real test kilometers required to validate HAD is too huge to be performed in a costand time-efficient manner [5], [6], [7]. Therefore, virtual assessment methods are needed in order to make validation processes more practicable [8], [9], [10], [11]. To reduce both time and risk, it is essential to have a realistic simulation environment for developing and assessing such systems in the early design phase [8].
The deployment of ADAS and HAD, which always implies safety issues, is decided by one central question: Will the traffic be safer overall for all road users by introducing the new system? To concretize this question, it can be divided into two parts: 1. Can the system reduce the occurrence of critical situations and/or accidents or mitigate their consequences? 2. Does the system induce new critical situations or accidents? Virtual traffic simulations can help answer these questions in the development process. They offer the possibility to examine new ADAS or HAD within realistic virtual test fields. The scope of virtual traffic simulations is broad. It is possible to assess overall systems, subsystems like sensors, and functional algorithms via hardware-in-the-loop (HiL), software-in-theloop (SiL), or Model-in-the-loop (MiL) simulations (see [10], [12], [13], [14], [15], [16]). The focus of virtual traffic simulations with detailed behavior models lies on the interaction between humans and the system to be validated. Simulations of future traffic with and without the system can be carried out and compared. This makes it possible to derive prognoses concerning the safety-related impact of the new technologies (see [17], [18]). In addition, evaluations regarding efficiency and comfort are possible. Traffic simulations that create realistic surrounding traffic are the prerequisite for virtual prognoses (see [19], [20]). Especially for safety evaluations, the simulation must represent realistic human behavior that leads to critical situations and accidents. In particular, the realistic simulation of junction scenarios is challenging due to the number of different road users and the complex processes a driver has to perform. In this context, realistic behavior means traceable This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/ actions based on cognitive processing (see [21]). These include the processes of perception, memory, decision making, and action implementation. For simple driving tasks like following, models which consider cognitive aspects already exist (see [22], [23]). However, there are no traffic simulations that meet these requirements for more complex scenarios such as urban junctions with multiple road users [24].
Another problem is that theoretically, a new system must pass an infinite number of possible situations to prove its safety. Many approaches have been developed to solve this problem (see [6], [25]). One approach is to simulate representative samples of situations with and without the system through a Monte Carlo study [6], [25], [26], [27], [28]. Therefore, performant simulations are needed that are capable of simulating a huge number of single simulation runs in a time-efficient manner. The demand for fast computing simulations that simultaneously guarantee a high degree of realism requires more efficient and realistic driver behavior models. This requirement is identified repeatedly by state-of-the-art reviews over the last few decades [24], [29], [30], [31], [32], [33], [34], [35], [36], [37]. Plavsic [24] states that the reviewed models either lag in representation of human cognition or are not freely available and extensible (see also [32], [37]). Especially for safety simulations, models need to be refined to better represent critical situations and accidents [29], [30], [31], [34], [35], [36]. The models are not able to represent the pre-crash phase, including the cognitive functions that lead to an accident. Mai [32] and Ni [33] note that most models do not reflect human heterogeneity.
This paper addresses the problems described above by providing the theoretical knowledge needed to develop more realistic driver behavior models. The paper focuses on the current blind spot in critical situation and accident modelling and provides a perspective on how the findings of human cognition could be used in behavior modelling. What makes the paper unique is the connection of in-depth knowledge about human cognition and the application of it. The structure of the paper is shown in Fig. 1.
Section II addresses what realistic traffic means and gives an insight into the causality of accidents. Subsequently, the causes and mechanisms a driver behavior model must consider in order to generate realistic behavior are presented (section III). Section IV deals with driver behavior modelling, explaining general concepts and the underlying modules. Current cognitive models and their issues are presented. As an alternative, the driver behavior model DReaM is introduced. At the end of the section, it is shown how the theoretical knowledge from Rasmussen's error taxonomy can be applied in driver behavior modelling. Subsequently, section V highlights the advantages of realistic traffic simulations. Finally, section VI summarizes the results.

II. WHAT DOES REALISTIC TRAFFIC MEAN IN SIMULATION CONTEXT?
The overall goal of driver behavior models is to simulate realistic traffic. In general, realistic traffic incorporates all possible behavior of a considered population, which influences the traffic within a certain area. This includes normal driving, critical situations, and road accidents. These categories are defined as follows: • A road accident occurs when at least one road user is injured or property damage is caused • Normal traffic is free of accidents • On the spectrum between normal traffic and road accidents, critical situations exist, in which the probability of road accidents is increased The change from normal to critical situations is characterized by the transition of accident probability from an unlikely to an increased probability of occurrence. Intended and unintended violations of traffic rules imply higher accident probabilities and therefore belong to critical situations. As criticality measures, time-based values such as time to collision (TTC), time headway (THW), or post encroachment time (PET) are often used. In addition to these, there are other surrogate safety measures (see [38], [39], [40]). When looking at critical situations, metrics that estimate the effort required to avoid accidents are of particular interest. Mentioned here are the steering threat number (STN) and the brake threat number (BTN). Both numbers express the ratio between the necessary acceleration (in lateral or longitudinal direction) to avoid a collision and the maximum available acceleration [41].
Risk, as the combination of the occurrence probability and the severity of a prospective event, is also used to express criticality (see [42], [43]). A maximum critical situation traverses into an accident event. Critical situations represent the linkage between normal situations and accidents. In order to fulfil the goal of realistic traffic simulations, driver behavior models must be able to simulate human behavior in normal driving, critical situations, and accidents. Therefore, the models must be able to explain the mechanisms that cause the back-and-forth transitions through these three phases because human drivers are also capable of preventing accidents through adequate avoidance actions in critical situations, resolving them back to normal driving.

A. Accident Causality (Where Do Accidents Originate?)
In order to simulate the transition from normal driving to accidents, the reasons for accidents in general have to be considered first. Critical situations and accidents are attributed to vehicular, environmental, and human error. Human error has the biggest influence on accident figures. More than 70% of all road accidents are due to human failures [44], [45], [46], [47], [48], [49]. Treat et al. [47] report that human errors and deficiencies were probable causes in 92.6% of 2,258 investigated accidents. Singh [49] relates 94% of all investigated accidents to human error. Dingus et al. [48] state that human-related factors were involved in 87.7% of all injurious and property damage events within the Second Strategic Highway Research Program Naturalistic Driving Study (SHRP2 NDS).
However, it has to be mentioned that not all human errors lead to accidents. Human errors imply the possibility of an accident, whereby different error categories individually increase the probability of an accident [48]. Humans make errors frequently but usually without serious consequences. The occurrence of accidents is relatively rare compared to human errors [44]. Moreover, errors can be corrected over time while remaining errors are accumulating. Accidents only occur when errors persist over time or when the time spans in which the driver can react are too short.
Another important point is that not all accidents are avoidable. There are accidents based on structural errors that happen even if all opponents react perfectly at the earliest possible moment [50]. Finally, participation in road traffic requires mutual trust. Road users cannot behave as if the others were permanently violating the traffic rules.

III. HUMAN ERRORS
This section explains the causes and mechanisms of human error. For this, possible error categories are presented first. Subsequently, Rasmussen's error taxonomy is applied to the traffic context and enriched by further research.
According to Reason [51], human error "[…] will be taken as a generic term to encompass all those occasions in which a planned sequence of mental or physical activities fails to achieve its intended outcome, and when these failures cannot be attributed to the intervention of some chance agency." Rasmussen [52] breaks human errors down to human-task misfits.
The identification of human error is not simple. It is especially difficult to determine the exact moment when the error causing an accident occurs as errors develop over time. For example, in a junction situation, where the driver has overlooked a pedestrian, when does the error occur? Is it the earliest moment the driver could physically see the road user or the moment the driver turned and missed looking over their shoulder?

A. Error Categories (Where Do Errors Occur?)
Subsequently, two human error taxonomies are presented. The first is derived from accident research and the second from cognition theories. According to Neisser, "[…] 'cognition' refers to all the processes by which the sensory input is transformed, reduced, elaborated, stored, recovered, and used" [53, pp. 4].
1) Taxonomy According to Accident Analysis: Treat et al. [47] defines five categories of human errors derived from externally observable causes in accidents: recognition, decision, performance, critical non-performance, and non-accident. Recognition errors include perception and comprehension problems. Decision errors refer to wrong chosen actions. Performance errors refer to insufficiently implemented actions that were correctly selected. The two last categories can be neglected. Critical non-performance (blackout or falling asleep) occurred in less than 3% of accidents in Treat's investigations. Non-accidents or intentionally caused accidents did not happen at all. Singh [49] uses the same categories and confirms Treat [47] that recognition and decision errors are the predominant human-related direct accident causes. Both studies attribute about 40% to recognition errors and 30% to decision errors over all human-caused accidents. According to Rumar [45], delayed detection is the main accident cause. Therefore, he proposes two classes: cognitive detection and perceptual detection errors.
Hakamies-Blomqvist [54] categorizes errors according to incapable of action, observation, estimation, and driving. The categories of observation, driving, and incapable of action are similar to Treat's recognition, performance, and nonperformance errors. Estimation errors refer to an incorrect evaluation of one's own or other vehicle's kinematic values or predicted decisions.
2) Taxonomy According to Cognition Theories: Reason [51] defines a general categorization of human error, dividing errors into three categories: slips, lapses and mistakes. Slips relate to failures in attention. Lapses involve failures of memory. Mistakes denote errors in the judgement and selection of objectives or steps to achieve them. Rasmussen [52] developed a sequential malfunction algorithm to identify the human mental function of the decision-making process. O'Hare et al. [50] adapted the model slightly to obtain the following categories: structural, information, diagnostic, goal, strategy, procedural, and action errors. Gründl [55] applied this categorization to 312 accidents. At 77%, recognition errors were by far the most common cause of the investigated accidents.
Almost all the error categories of the two taxonomies originate from human cognition mechanisms. Wagenaar et al. [56] support this idea and argue that in an industrial context, human cognition is the most common cause of accidents.

B. Causes of Errors (Why Do Errors Emerge?)
After focusing on human error categories, the emergence process of human errors is considered. In general, human errors are due to the human information process or the individual's mental model [44], [57]. Human errors build on internal mechanisms, which will be triggered by error causes. For this  Table I). Rasmussen lists the following major causes of human error: • External event (distraction) • Excessive task demand • Operator incapacitated • Intrinsic human variability 1) External Event (Distraction): Human cognition is limited in processing information. Therefore, attention must be directed to task-relevant information. Tian et al. [58] found a connection between off-road eye-glance duration and the probability of crash and near-crash events. The process of filtering for relevant information while ignoring other information is called selective attention [59]. Combined with events that distract attention, important information can be missed, leading to accidents [60]. In this way, sudden onset visual stimuli can attract attention [61]. Furthermore, other external events or activities can disrupt the driver. Cell phone use while driving is a striking example of disrupting performance. Even for hands-free devices, Strayer et al. [62], [63] suspect that drivers are distracted due to the distracting effect of the conversation. Dingus et al. [48] deliver astonishing figures for distraction. They show that distraction was observable in 68.3% of investigated accidents. In addition, over 50% of the time, some kind of distraction prevents drivers from focusing on the main task of driving.
2) Excessive Task Demand: The degree of task demand is of particular importance in the transportation context. Based on the underlying scenario, the task demand can vary widely. If the task demand exceeds an individual's capability, they will be unable to perform the task in the required manner [64]. Task demand represents the complexity of the situation. It includes all values the driver has to consider according to their goal, including the behavior of other road users, environment, kinematic and spatial parameters, and especially the driver's speed. Fuller [64] defines capability as the "[…] momentary ability of the driver to deliver his or her level of competence", where competence is characterized by the three levels of performance [65] and is affected by human factors like fatigue, motivation, or influence of drugs or alcohol. The task difficulty can be expressed based on the variables of task demand and capability. Where task demand exceeds the capability, the task is too difficult [66]. Fuller [66] demonstrates how excessive demand can successively become visible, starting with the fragmented degradation of tasks, then leaving out low-priority tasks, up to skipping high-priority tasks.
Next to task difficulty, workload is also an indication of excessive demand. Workload is strongly connected to task difficulty. "On the basis of task difficulty processing resources are allocated and mental workload is reflected by the amount of allocated resources" [67]. Reimer [68] suggests that visual attention is affected by the workload of secondary tasks. He notes that by increasing workload, gaze restriction occurs before vehicle control suffers. Brookhuis and de Waard [69] state that the appropriate amount of mental workload is important for accident-free performance. Too little workload can affect alertness, just as too much workload can cause stress. Both increase the probability of accidents occurring [67], [69].
Stress is a nonspecific body response to a combination of external demands and internal concerns [70]. Stress causes a high level of mental workload and can affect the perception process this way [70]. Stress can limit people's ability to process stimuli by affecting signals that initially attract less attention [71]. By the high level of workload, stress contributes to the process of task degradation (see [66]). In addition, feeling stressed is a contributing factor to crashes [72]. Stress can influence the driver's mode (anger or aggressiveness) and lead to riskier driving behavior [70].
3) Operator Incapacitated: This category includes driver impairments such as sickness, drugs/alcohol, drowsiness/ fatigue, or emotion that adversely affect driving. Dingus et al. [48] examined the SHRP2 NDS data according to the influence of these driver impairments, with the exception of sickness. Overall, 1.92% of baselines showed driver impairment, which increased the risk of an accident 5.2 times compared to normal driving. Mortazavi et al. [73] found a significant correlation between drowsiness and the degradation of driver's lane keeping and steering ability. Driving under the observable influence of drugs or alcohol increases the risk of an accident 35.9 times. Drowsiness and fatigue was observed in 1.57% of baselines and increased the risk of an accident 3.4 times. The analyses according to emotional state -anger, sadness, crying -are interesting [48]: Driving in such an emotional state increases accident risk 9.8 times.
4) Intrinsic Human Variability: Intrinsic human variability refers to the fact that humans do not react exactly the same way in the same situation. Intrinsic human variability already sets in by processing simple stimuli. Mayhew et al. [74] showed this for the brain and behavior response to pain. The studies of Fox et al. [75] suggest that intrinsic brain activity correlates with variability in human behavior. They conclude "[…] that inconsistencies in perception or performance should not automatically be attributed to fluctuations in task-related cognitive processes such as attention, but could also be due to persistent fluctuations in intrinsic neural activity" [75]. Currently, the intrinsic variability for behavior is poorly understood.
Human variability is also reflected in driving style. Taubman-Ben-Ari et al. [76] suggest eight main factors influencing driving, each one expressing a specific driving style: dissociative, anxious, risky, angry, high-velocity, distress reduction, patient, and careful. Based on these factors, Taubman-Ben-Ari et al. were able to predict self-reports of involvements in car accidents and the occurrence of driving offenses.
The individual driving behavior is based on several factors, like individual factors, sociocultural aspects, and technological factors [77]. Individual factors are gender, age, experience, or personality traits [77]. The results of Poó and Ledesma [78] indicate that individual differences in driving styles can be explained due to individual personality traits. Lucid et al. [79] showed that accident risk correlates with age. The research of Taubman-Ben-Ari et al. [76] indicates that personality traits such as sensation seeking contributes to accident involvement.
In addition to questionnaires (see [76], [80], [81]), objective methods based on recorded driving data (see [82], [83], [84], [85]) are also used to determine driver behavior. Driving styles, which are characterized by underlying measured kinematic values, can be directly applied to driver behavior models by using these values (acceleration profiles, reaction times, speed ranges) for the parameterization. Moreover, recoded personality traits can be used to individualize ADAS functions to enhance the acceptance of the function (see [86]).

C. Error Mechanisms (How Do Errors Occur?)
After looking at why humans fail, this section focuses on how errors occur. For this purpose, Rasmussen [52], [87] developed a taxonomy of mechanisms of human malfunction. This section applies Rasmussen's taxonomy to the traffic context and enriches the categories with further research. Rasmussen suggests five main categories of human error mechanisms: • Discrimination • Input information processing • Recall • Inference • Physical coordination 1) Discrimination: Discrimination includes all failures related to biases or fixations in the choice of interpretation of a given situation and the conclusion drawn from it [52]. Humans make decisions based on their mental models. Mental models [46], [88], [89], [90] contain the whole set of cognitive functions needed to build a representation of the external reality. The mental model stores all perceived information and associated interpretations in the cognitive map [91], [92], [93]. The cognitive map contains a personal point of view, including failures and incorrectness. Based on their mental models, humans analyze the situation, predict the future, make plans, and derive actions.
If the underlying situation is not compatible with the mental model, errors may occur. This happens when situations are new or rare. To learn and optimize suitable processing patterns, an individual needs opportunities to perform trial-and-error experiments [52]. Even if situations seems to be familiar and suitable processing patterns are available, errors are not excluded. Based on the vast number of situation parameters involved and their variability, situations can develop unexpectedly.
Rumar [45] illustrates this construct for his cognitive detection errors category. He notes that perceptual analysis must be highly selective due to the enormous richness of visual stimuli. That leads to specific search patterns to investigate the most important areas according to the mental model. According to Rumar [45], vision works as a testing instrument for hypotheses formulated by the mental model. If the search patterns do not match the situation, important information can be missed (e.g. the tendency to overlook cyclists coming from the right when turning right on non-signalized junctions in right-hand traffic; [94]). Summala et al. [94] explained this behavior by hypothesizing that people develop scanning patterns towards more likely hazards, with the risk of not considering rarely occurring hazards. Rasmussen [52] called this stereotype fixation. In the example above, the driver does not expect the cyclist to come from the right, because riding on the wrong side of the road is against traffic rules, and road users do not expect other road users to violate traffic rules willingly by default. Besides the visual search pattern, this applies to all mental processes like extrapolation or action determination. Reason [51] states that humans are strongly disposed to stick to already available patterns (see also [95]). He even designates humans as "[…] furious pattern matchers." 2) Input Information Processing: This category is closely related to discrimination because an incorrect information base leads directly to the false selection of patterns. Rasmussen [52] associates not perceived or wrong processed information with this mechanism. The visually perceived information is the most important source of information for the mental model. According to Sivak [96], it is often reported that the majority of information needed for driving comes from visual input. Rumar [45] states that human functional perception is a product of ecological adaption and is not made for the artificial environment of traffic. Therefore, in certain situations, important stimuli, such as other road users, are below the perception threshold and will not be processed.
Information processing is strongly related to attention. Human perception and mental resources are finite. Perceiving the "right" information and processing the "right" tasks to remain accident-free depends on where our attention is focused (see [97]). Wickens and McCarley [98] suggest that attention should be understood as a front-end filter that only lets pass certain stimuli or events for processing, depending on the underlying task. Afterwards, the downstream mental resource limits the number of processes that can be performed [98]. Simons and Chabris [99] showed how important the focus of attention is for what humans really perceive. Without attention, people often fail to recognize major changes in objects or scenes (change blindness; see also [100]) or even fail to notice conspicuous objects if their attention is on another object or task (inattentional blindness; see also [101]). Mental processes and inner assumptions control attention, which can replace the search for information in task-relevant areas and direct attention to less important areas (see [87]).
According to Reason [51], errors often depend on failures of attention, more explicit due to inattention by missing task-relevant information or overattention by disrupting automatized action sequences by attentional checks. Therefore, Reason [51] counts these failures of attention as control mode failures.
3) Recall: This category includes the mechanism that relates to human memory. Humans forget data. Moreover, a wrong action can be chosen among various alternatives. The literature distinguishes between short-term, long-term, and working memory (see [102]). Long-term memory contains knowledge in the form of semantic networks, episodic memory schemata, and mental models [103]. In the context of traffic, these are traffic rules, behavior rules, driving skills, or experience. Rasmussen mentions that when recalling mental models, isolated acts can be forgotten, which results in errors [87].
The idea of working memory evolved out of the concept of short-term memory [104]. Besides storage, it also processes and manipulates the perceived data. The capacity of working memory is limited and contains temporal decay [102], [105]. The perceived information is interpreted and combined with long-term memory knowledge. This active knowledge about the underlying dynamic scene is called situation awareness [106].
Most of the time, the information base on which decisions are made is incomplete due to constraints in perceptual processing, working memory, long-term memory, or external limitations. For example, long-term memory fills information "gaps" through experience, which can lead to inappropriate rules being triggered due to the added information [108].
Another cause for failures is the limited attention capacity. Direct attention is not only reserved for perception tasks; it is also needed in decision-making and response execution [103]. Thereby, the concentration of attention on one task can interfere with other tasks [108].

4) Inference:
This category relates to failures in interpretations and predictions of the given information. Based on the situation awareness, schemata or mental models of the longterm memory are retrieved and applied. Failures in maintaining situation awareness often lead to errors in real-time tasks like driving or flying [106].
The working memory must also predict the future and draw conclusions from it in order to select appropriate actions to maintain the current status [103]. Errors may occur if the prediction is incorrect or side effects are overlooked. Unexpected lane changes are striking examples. While approaching junctions or overtaking on highways, the driver does not expect sudden lane changes from other road users. Here, the driver hypothesizes that other road users know his or her position and will not move into his or her lane. Therefore, drivers extrapolate the lane changer as a "lane follower". This conclusion could lead to accidents if the driver is not ready to break for an upcoming lane changer. The fact that the lane changer is the main causer does not negate the fact that the driver's mental model is wrong, and he or she chooses an incorrect pattern. 5) Physical Coordination: Rasmussen [87] assigns all errors based on insufficient execution to this category. The underlying goal and the chosen action are appropriate, but the person does not implement the action properly. Thus, the individual does not perform as they aim for due to motor variability or spatial orientation failures.

IV. DRIVER BEHAVIOR MODELING
There are two fundamental directions in driver behavior modelling [109], [110]: 1. Predictive models 2. Cognitive (explicative) models Predictive models are focused on the reproduction of the behavior itself. They generate behavior based on pre-observed data. These models are descriptive. The developers know what people will potentially do, but not why [111]. These models do not consider the underlying processes that lead to the behavior.
Cognitive models try to model the internal processes and states that produce the behavior. The driver's decisions and actions do not depend directly on the environmental state but rather the mental representation and interpretation made out of it [109]. In this way, they are adaptable to changing situations as long as the required cognitive processes remain constant.
Both model approaches have their justification. Predictive models can help to simulate behavior when the underlying cognitive processes are too complex, not understood well enough, or the computing time is too intensive. The disadvantage is that, especially for critical and accident situations, it is difficult to create such a model. Human errors arise and accumulate over time. Therefore, it is difficult to establish a relationship between the data basis and the accident-causing behavior without considering the processes of cognition. These models are limited in their predictive capabilities due to noise. The noise is based on the error causes and mechanisms from section III, which are not directly mapped by predicted models. In addition, the data basis of critical situations and accidents is often small and/or of low level of detail.
Cognitive models could help to close this gap. Most of the time, critical situations and accidents arise based on cognitive processes. By implementing these cognitive processes, errors can be simulated based on internal processes such as incomplete information (looking in the right direction but not perceiving important information like the collision opponent), incorrect extrapolation of information (incorrect prediction of the movement of the collision opponent), or misinterpretation (choosing a wrong action). In addition, cognitive models are adaptable to new situations as long as the new situations are based on the same cognitive processes. This characteristic is essential for the virtual assessment of new systems that potentially induce new situations. Without this model characteristic, road users would not react realistically in new situations, and predictions would not be possible.
Predictive models and cognitive models are not necessarily distinct. Computational cognitive driver models often have predictive sub-models, if cognitive functions are too complex or not yet understood. Steering is often modelled by a simple proportional-derivative (PD) or a proportional-integralderivative (PID) controller (see [112], [113]). Machine learning approaches gain weight in trajectory planning [114], [115]. So too are car-following models are mostly based on kinematic values like time headway (see [116]) instead of retinal projections of the leading vehicle or the visual flow field around the leading vehicle [117]. Basak et al. [118] stated that reaction time models often do not match actual human behavior. Another example is the modelling of gaze behavior based on pre-observed gaze data (see [4]).

A. Driver Behavior Modules
This paragraph gives a more detailed inside into driver behavior modules and algorithms to achieve realistic driver behavior. The focus is on microscopic models, capable of simulating the interactions of individual road users (agents). Widely-used traffic simulations like PTV Vissim (see [119], [120], [121]), SUMO (see [122], [123]), PELOPS (see [24]), or IPG Carmaker (see [124]), are able to simulate traffic flow, but with limitations on the insight of driver behavior modeling. Studies that use these traffic simulations to model critical and accident situations are (so far) tests of feasibility rather than actual simulations of real-world conflict situations (see [35], [125], [126], [127], [128]). Subsequently, a classification of currently used modules in driver behavior models is presented. (see Table II). As mentioned in section II, realistic traffic covers normal traffic, critical situations, and accident situations. Along this division, the driver behavior modules are explained.
1) Normal Traffic: Three different modules must be included to represent normal traffic flow: a car-following model, a lane change model, and a traffic infrastructure node model. They lay the foundation for longitudinal and lateral guidance and hand over the desired kinematic values of the agent to the traffic simulation. All driver behavior models rely on this functionalities.
Car-following models are historically the oldest parts of traffic simulation. Deterministic car-following models calculate the acceleration of the ego-agent to reach the desired speed that depends on the distance and/or time headway to the leading vehicle and is based on kinematic equations. The most popular deterministic car-following models are the Optimal Velocity Model (OVM) [129], the Gipps Model [130], the Generalized Force Model (GFM) [131], and the Intelligent Driver Model (IDM) [132]. Next to these deterministic models, also psycho-physical models exist, such as the Wiedemann Model [133]. The ego-agent adjusts its relative distance and relative velocity to the leading vehicle according to a minimal and maximal threshold (see [134], [117]). Panwai and Dia [135] assess the accuracy of car-following models implemented in the following traffic simulations: VISSIM (Wiedemann Model), PARAMICS (Wiedemann Model), and AIMSUN (Gipps Model). Machine learning models are also used to model car-following (see [136], [137], [138], [139]). They map traffic flow well, but have drawbacks due to their predictive model nature (see above). There are also combinations of kinematic car-following models and machine learning approaches. The Combination Car-Following Model (CCF) of Yang et al. [140] combines the result of a machine learning model and the Gipps Model by a weighted value.
Lane changes are necessary to navigate to a specific destination or select a lane that better matches the desired speed. Therefore, lane change models have evolved from carfollowing models that already control speed in the current lane. Lane change models typically consist of a strategic, a tactical, and an operational stage [141]. In the strategic stage, the driver chooses a new lane according to the state of the current lane (blocked, lane speed, destination). In the tactical stage, the upcoming lane change is prepared by acceleration or deceleration. Subsequently, it is determined that the lane change is safe and desirable [142]. The decision process is usually modeled as a binary gap problem [143]. Parameters such as urgency and risk affect the selected gap. According to Rahman et al. [144] lane change models can be categorized in four groups: rule-based models (see [142], [145]), discretechoice-based models (see [143], [146]), artificial intelligence models (see [147], [148]), and incentive-based models (see [141], [149]). Rule-based models are based on decision trees with fixed conditions. Discrete-choice-based models utilize logit or probit models. Artificial intelligence models are trained by field-collected data and incentive-based models decide on lane-change desire [144].
Most traffic simulations are based on agent's modules that follow a reference lane. Next to this, also trajectory planning algorithms exist, according to specific driver objectives (e.g. comfort, lane keeping, minimized speed variation) (see [150]).
Traffic infrastructure node models cover all algorithms to simulate traffic at infrastructure nodes that connect different roads. The main tasks are determining the priority, predicting the current intersection gap, and adjusting the acceleration/deceleration to take or skip the gap. Determining priority can be done simply by applying the traffic rules. Doniec et al. [151] express the priority status of each agent over another by a Boolean. In addition, they implement an algorithm to give up priority if a specific agent blocks a junction to avoid deadlocks. If an agent does not have priority, the current gap with the conflict-agent is evaluated. The ego-agent can go if it passes the conflict area before the conflictagent reaches the conflict area or vice versa. Otherwise, the speed must be adjusted (see [151], [152], [153], [154]). In contrast to gap acceptance models based on kinematic equations, there are also machine learning approaches (see [155], [156]).
2) Critical and accident situations: Based on normal traffic, the mechanisms of human cognition have to be included to simulate critical and accident situations. Several approaches exist to simulate critical and accident behavior. The most common are: reaction time models, perception models, and cognitive models. These models form an additional cognitive layer within driver behavior models.
Reaction time models simulate the human time delay from the perception of a stimulus to the execution of an action. There are different reaction time studies and definitions for a more detailed insight (see [157], [158], [159], [160]). In the beginning, reaction time models were used in traffic flow simulations to reduce velocity oscillations in dense traffic and represent a more natural traffic flow (see [161], [162]).
However, reaction time also has a crucial influence on accident figures, especially in the case of sudden or unexpected movements of other road users or the appearance of road users form behind visual obstructions. There are different approaches to implement reaction time. Many models use fixed reaction times [129], [130], [162]. Basak et al. [118] state that response time modelling is limited in most models and is more an artefact of the underlying simulation step size. Therefore, Basak et al. [118] developed a more sophisticated framework for reaction time modulation. It enables the selection of different independent variables (e.g. speed, distance or indicator state). For each selected variable, a reaction time is drawn based on the underlying distribution. The continuously perceived values are buffered and depending on the reaction time, the corresponding value from the buffered array is forwarded.
Lindorfer et al. [127] used almost the same approach, where instead of an individual selection of values, a fixed set of values (ego speed, net distance, velocity difference) is delayed. Finally, some models attempt to model distraction by adding a stochastically drawn brake reaction time (see [163]).
Perception models simulate the acquisition of environmental information. Visual information is the most important source for driving [96], which is why almost all perception models consider only this sensory channel. The simplest way to simulate the imperfect human perception process is to perceive all surrounding information and add noise.
Noise Models aim to reproduce the human error by adding a variance (noise) to a value, e.g., a stochastically generated error to the perceived distance, speed, or acceleration (see [128], [162]). This contradicts Lee's τ -theory [107], according to which humans use the change in the optical image at the driver's retina for TTC determination instead of using speed, distance, and acceleration estimates for longitudinal guidance.
Another approach to provoke critical situations is to simulate driver distraction by blocking the acquisition of new environmental information for a certain time (see [58], [125], [127]). More advanced models try to simulate the actual useful field of view (UFOV) [164]. The UFOV is represented as a two-or three-dimensional funnel, where information (speed, acceleration, position) about the surrounding agents can be perceived (see [116], [165], [166]). In this way, the ego agent has only up-to-date information about the observed area. In addition, endogenous and exogenous visual attention [167], also known as "top-down" and "bottom-up", can be implemented by splitting the perception process into a knowledge based top-down and a saliency-driven bottom-up process. The problem with this model type is that the opening angle and gaze behavior (where the visual funnel is directed) are difficult to parameterize. To parametrize these top-down and bottomup processes, Denk et al. [21] and Horrey et al. [168] suggest using the SEEV model (salience, effort, expectancy, value). The idea behind the SEEV model is to predict the probability of attention (eye fixation) based on the perceived parameters of saliency, effort, expectancy, and value [98]. It is also conceivable to map gaze behavior using machine learning, but no such models exist yet.
The introduced models are often coupled. Treiber et al. [162] modeled estimation errors for distance and velocity stochastically and showed that destabilization effects (estimation errors, reaction times) can be compensated by anticipation of other road users in order to simulate congested traffic with realistic reaction times. Reaction time and perception are single cognitive aspects of human information processing. In this context, cognitive models refer to broader modeling approaches that incorporate the entire cognition process. This models include: perception, memory, decision, and action implementation (see [169], [170], [171]). There are cognitive models with different origins. Driver behavior models based on existing psychological cognitive architectures, such as ACT-R [172], Soar [173], [174], or QN-MHP [175], have the advantage that they should be psychologically valid; but they are also very complex [24]. ACT-R and Soar have a high level of abstraction, slow runtimes, and are not adaptable for arbitrary dynamic environments, making them unsuitable for complex scenarios like behavior modelling for urban areas [24]. The cognitive architectures ACT-R and QN-MHP were used in several studies to quantify aspects of driver cognition, such as lateral and longitudinal guidance, reaction time, distraction, or speed control (see [23], [176], [177], [178], [179], [180]). So far, these models have been used to study driver behavior and not to simulate critical situations or accidents. Due to their complexity and their structure, which is not designed for performance, they are not suitable for large simulation studies such as Monte Carlo studies.
COSMODRIVE [109] is a pure driver behavior model. The perception process is implemented by means of a "virtual eye". It is able to simulate data-driven (bottom-up) perception of information according to their saliency. In addition, topdown perception is also modeled, where fixation is driven by the driver's plan [181]. The perceived information is matched to existing driving knowledge (driving schemata and envelop zones) to determine future actions that meet the tactical goal of the activated driving schema [182]. In combination with Pro-SiVIC, a vehicle-environment-sensors platform, COSMOD-RIVE was used to develop a Virtual Human Centered Design (V-HCD) platform [181], [183]. V-HCD is potentially a perfect candidate for ADAS or HAD testing, due to the driver modelling on one side and the realistic sensor-vehicle simulation on the other. Unfortunately, the software is not open-source. In addition, only very simple simulation scenarios were presented to the public. ACME Driver [184] was developed to extend the traffic simulation SUMO for microscopic driver modelling. The general structure is divided into a sensor unit, a processing unit and an action implementation unit. In addition, a sensor filter and action implementation filter exist. Unfortunately, no more detailed information about the model can be found.
SSDRIVE [185], [186] consists of the components environment, driver, and vehicle. The resulting driving behavior depends on (1) a set of simple rules, based on task analysis; (2) five parameters, influencing the interaction with the vehicle-environment for different driver types; (3) mechanisms for decision making and error generation. The five parameters (experience/competence, attitudes/personality, task demand, driver state, situation awareness/alertness) are based on driver aspects (driven kilometers per year or years of driving) and the driver-vehicle-environment interaction (traffic complexity, road conditions, distraction, speed) [186]. However, these correlations are based solely on expert knowledge and are supported by limited experimental validity [185]. The errorgeneration process depends on the Driver Impairment Level (DIL), which is calculated based on all five parameters. When the DIL is equal to one, an error occurs. The type of error is determined by the user, according to the underlying action. The modelling approach of SSDRIVE is interesting, but there is no relevant further development since 2010. The model is assumed more of a prototype and it is not available for free usage.
The SCM is especially developed for the prospective safety assessment of ADAS and HAD. Unlike other models, it is able to simulate human variability by using probability distributions to determine driver behavior parameters and driver actions. SCM consists of six submodules: gaze control, information acquisition, mental model, situation manager and action manager (decision making process), and action implementation [19]. Gaze control and information acquisition include the peripheral and foveal UFOV as well as the modelling of top-down and bottom-up perception. In the mental model, all information (current perceived information, working memory information, and information from previous time steps) is aggregated and processed. The situation manager and action manager assess the current situation based on the mental model information and determine the next driver action. Finally, the chosen action is implemented by adjusting the pedal positions and the steering wheel angle. Up to now, SCM is only able to simulate highway traffic. Furthermore, SCM is not validated and the source code is not freely available.

B. DReaM -a Driver Behavior Model
Subsequently, the structure of the cognitive model DReaM 1 (Driver Reaction Model) is explained (see Fig. 2), which is developed by Christian Siebke. DReaM supports the current open traffic simulation formats openDRIVE [187] and openSCENARIO [188]. The focus of DReaM is on urban traffic, especially on junction scenarios (see [165]). The used parameters are taken from stochastic distributions. This way, realistic traffic is obtained by simulating the behavior of a random sample of individual road users. The general structure of DReaM is similar to other cognitive models and is divided into perception, memory, decision, and action implementation [169], [170], [171].
Much effort is spent on the Perception process. The UFOV is modeled as a two-dimensional swiveling funnel. Similar to COSMODRIVE, the bottom-up and top-down perception process is modelled. The bottom-up fixation is parametrized by a simulator study (see [189], [190], [191]). Top-down gazes are trigged, e.g., by turning maneuvers to watch for pedestrians/cyclists or to observe agents with whom a collision is imminent. Inside the Working Memory, the perceived information is stored in the Cognitive Map. The capacity of the working memory is limited. This way, only a maximum number (user adjustable) of processed agents can be stored in the cognitive map.
In each time step, the perceived information is buffered in a data array. According to the reaction time, the information is accessible x time steps delayed (see [118], [127]). There are two reaction times, which can be parametrized individually: initial reaction time and latency. The initial reaction time represents the process of a target appearing for the first time. Here, the driver is unprepared and surprised. Therefore, the reaction time is often extended. The latency describes the periodic perception delay, e.g., when observing a target. 1 Available at: driver-model.de. The kinematic information of agents not updated by a gaze fixation is extrapolated based on available data from the cognitive map. In addition, the aggregated information is interpreted and specific situation states are drawn. The Interpretation is based on schemata or mental models of the Long-Term Memory, which are implied by the code. Situation states are right-of-way, following, collision with another agent, current phase on the junction, or hazard (crash is imminent). All the information combined forms the ego-agent's situation awareness, the basis for decision-making.
The Decision process is divided into Gaze Movement, Lateral Decision, and Longitudinal Decision. The bottom-up gaze movement is parametrized by probability distributions via the central behavior configuration file, where all used DReaM parameters can be adjusted. The top-down gazes are derived from the performed maneuver, the geometry of the junction, and the current situation. While approaching a junction, important fixation points are determined. The first kind of points are on the lanes leading to the junction. Here, other potential conflict agents must be observed to determine the right-of-way and to react if someone violates it. The second kind of points are at the corners of junctions, where pedestrians and cyclists cross the road. In addition, there are observation gazes, where other agents are fixated: First, if a collision is imminent and second, if the ego-agent follows a leading vehicle. How often each top-down gaze is triggered is also parametrized by a probability via the behavior configuration file [166].
The Lateral Decision selects the current lane according to the calculated graph (route to a target). DReaM calculates the graph during the initialization of an agent, based on the waypoints of the openSCENARIO file. If the agent needs to change lanes to stay on the graph, the module sets the indicator (if the adjacent lane is free) and calculates the lateral displacement to the new lane. In the Longitudinal Decision, the required acceleration is calculated to maintain a desired speed without colliding with surrounding agents. The acceleration calculation is based on the IDM [132] car-following model. The present situation with each agent is evaluated and the lowest acceleration/deceleration resulting from the evaluation is chosen. In the Action Implementation, the lateral and longitudinal action is implemented. In each time step, the lateral displacement and orientation error (difference between vehicle and lane orientation) to the reference lane is calculated. By means of a Stanley controller [192], the steering wheel angle is calculated to keep the agent on track. The longitudinal acceleration is passed to the longitudinal vehicle model, which has to limit the acceleration according to its physical boundaries. So far, a first rough validation of the resulting approach behavior at junctions was conducted [15]. In addition, the feasibility of assessing ADAS and HAD with DReaM was demonstrated, using an autonomous emergency brake (AEB) system for junctions (see [15]). Fig. 3 shows the change in accidents by implementing the AEB. On the left picture, the driver notices the cyclist too late and a crash occurs. On the right picture, the crash can be avoided by the reaction of the AEB, which detects the cyclist earlier due to a wider sensor opening angle.
DReaM is coded in C++ and designed for minimal runtimes under the condition of sufficient accuracy, to perform safety-relevant studies. Most driver behavior models are fitted for a specific scenario. However, real traffic behavior is often affected by local phenomena (visual obstruction, infrastructure, environmental influences).
DReaM should be seen more as a framework as solely a driver behavior model. Attention is paid to its modular architecture. All introduced modules are easily customizable and extendible. For example, certain situation states can be easily added or even entire modules can be replaced. In addition, all used parameters can be adjusted by a single behavior configuration file. This way, the driver behavior can be efficiently adapted to the investigation scenario.

C. Application of Human Cognition and Error in Driver Models
After getting a detailed insight into Rasmussen's error causes and mechanisms on the one hand and driver behavior modelling on the other hand, now the focus shifts on what can be done to address Rasmussen's taxonomy. This paragraph suggests how Rasmussen's error causes and mechanisms can be applied to obtain more realistic driver behavior models.
One starting point to obtain more realistic driver behavior is the mapping of human variability. This is of particular interest for virtual safety assessment, as a better match between the real and simulated driver populations increases the precision of safety predictions. Human variability includes intrinsic human variability as well as variability between different individuals. Both can be addressed by statistical implementation of the driver behavior parameters, like in SCM [4] or DReaM. To do so, reliable data is needed. There are approaches to identify more precise driver types/driving styles on recorded data (see section III.B.4). Based on the characterizing parameters (acceleration profiles, speed ranges, reaction times) of these driving styles, driver behavior models can be parametrized. Driving styles and personality traits are important aspects that contributes to realistic driving profiles, critical situations, and accidents. Simulated traffic becomes more realistic by mapping these individual aspects in driver behavior models.
The modelling of different driver states (distracted or impaired) like mentioned in sections III.B.1 and III.B.3, paired with related penetration rates, would contribute to fit the resulting virtual behavior to real behavior. Impaired (incapacitated) drivers can be modelled by higher reaction times, lower THWs, or divergent acceleration profiles. Lindorfer et al. [127] simulated distracted drivers by increased reaction times or by not updating the input parameters for a certain period of time, like in [125] or [58]. Furthermore, stimuli fixation can result in distraction. Therefore, bottom-up stimuli perception has to be implemented, which can be addressed by the SEEV model.
Next to this, the driver task demand plays an important role (see III.B.2). Excessive task demand results in driver errors, starting with the fragmented degradation of tasks, then leaving out low-priority tasks, up to skipping high-priority tasks [66]. This mechanism is expected to have an even greater impact on urban scenarios, such as junctions, where task demand is high. There are already existing driver models in cognitive architectures that consider task demand (e.g. ACT-R [23], Soar [174], QN-MHP [175], or SSDRIVE [186]). Depending on the architecture, the processing of cognitive functions can be serial, i.e., only one task can be processed at a time. This way, tasks can be skipped.
Modelling the input information process (see III.C.2) offers the potential to simulate accident types that can be attributed to this cognitive process. Especially the visual perception process is crucial, as it is the most important information source for human drivers [96]. All human actions result from the situation awareness, which is mainly build on the visual information process.
Almost no driver model simulates the actual perception process, including the UFOV (the viewing area where information can aggregated). Instead, agents have direct access to all vehicles and noise is subsequently added or kinematic values are not updated temporarily (see [58], [125], [127], [128], [162]). There are only a few models that simulate the UFOV (see [19], [116], [165], [166], [181]). Wickens and McCarley [98] theorize that attention can be modelled by the perceived parameters of saliency, effort, expectancy, and value of the fixation target. Denk et al. [21] and Horrey et al. [168] suggest the usage of the SEEV model [98] of Wickens and McCarley for the modelling of bottom-up and top-down perception processes in driver models. Next to this, Wang et al. [116] developed an approach to simulate the visual perception process based on the so called FLMP  [193]. It is also possible to map gaze behavior by using machine learning, but up to now, this approach has not been implemented.
In a next step, the perceived information has to be stored. As mentioned in section III.C.3, the capacity of working memory is limited, which leads to a loss of information. This can be easily implemented by reducing the storage capacity of the driver model, like in DReaM. Contrary to that, it is rather difficult to operationalize the storage process and to parameterize the models.
Further sources of critical situations and accidents can be the forgetting of isolated acts or the wrong choosing of actions/tasks among alternatives. In cognitive models, this decision process of actions is often based on rules. However, some models additionally allow a probability-based selection for specific tasks, like the next gaze state (see DReaM or SCM). This way, there is a certain probability of choosing the wrong action.
The foundation of discrimination mentioned in III.C.1 are Rasmussen's causes of error: distraction, excessive task demand, operator incapacitated, or human variability. These human malfunctions lead to an imperfect situation awareness, which leads to failures. In addition, human-like action patterns have to be implemented. 'Human-like' means 'how the driver would actually react, not how to best handle the situation'. Used patterns in cognitive models are mostly based on expert knowledge, not on studies or recorded data. Again, the SEEV [98] model could help to trigger these realistic, human-like gaze patterns. Moreover, further naturalistic driving studies and driving simulator studies are needed on how road users actually react and what patterns are part of their repertoire.
Furthermore, the resulting working memory data have to be extrapolated, based on their kinematic values and the predictions of the mental model. Based on the conclusions drawn from a correct assessed situation, errors may occur (see III.C.4). This is a main task of cognitive models presented in section IV.A.2. The general problems with existing cognitive models are their adaption or the use of single functionalities from them in such a way that they can also simulate complex scenarios. In addition, the maturity and universal use of the models is not given. All presented cognitive models are prototypes and most of them are not freely available for adaption or further development.
Finally, to address failures in physical coordination (see III.C.5), noise models can be used to add a variance (noise) to the outcomes (acceleration/deceleration or steering angle) of cognitive models. Table III summarizes how Rasmussen's error taxonomy can be applied by state-of-the-art models and approaches. The color code is a qualitative estimation on how well the current models represent the mechanisms.
V. BENEFITS OF REALISTIC TRAFFIC SIMULATIONS Road safety assessment will benefit greatly from the introduced developments. Implementing Rasmussen's [52] human error causes and mechanisms into driver models increases the number of accident types that can be represented virtually.
This enables the virtual safety assessments of safety systems addressing each of these accident types. In addition, the prediction capability of virtual assessment increases by enhancing the underlying processes of traffic simulations. Moreover, the simulated accidents become more transparent as they can be traced back to their cognitive processes. By identifying scenarios relating to the same cognitive process, an additional basis for the development of new ADAS can be established.
Realistic driver models can be used for the generation of realistic road user behavior within virtual test fields of MiL, SiL, or HiL simulations (see [194]). This way, sensors, functional algorithms, or overall systems can be assessed (see [10], [12], [13], [14], [15], [16]). Virtual assessment is ideal for sensitivity analysis of how individual parameters (e.g. the sensor opening angle) (see [15]) affect the system behavior and the resulting traffic. Therefore, MiL, SiL, and HiL contribute to cost and time saving, risk reductions, and safety benefits by finding new scenarios through extensive simulation studies. In addition, virtual tests are reproducible [195].
Especially for stochastic virtual testing (see [6], [25]), such realistic traffic models are needed to generate critical and accident situations [196]. In this way, safety systems can be tested against new scenarios resulting from the realistic behavior of surrounding road users.
Human-like automated driving can benefit from driver behavior models to better cope with complex scenarios [197]. ADAS or HAD can use realistic driver models for the prediction of surrounding road users' behavior, to guarantee a smooth collaboration between humans and automated vehicles [198], [199].

VI. CONCLUSION
There is a strong demand for new performant traffic simulations for the assessment of ADAS and HAD. These new traffic simulations have to be able to simulate realistic traffic. Realistic traffic covers the whole spectrum of traffic, from normal driving over critical situations to accidents. In order to be able to simulate critical situations and accidents, the cognitive mechanisms of the individual that lead to behavior must be represented in the model. Human error is the main cause of road accidents. More than 70% of all road accidents are based on human error [44], [45], [46], [47], [48], [49]. In line with Rasmussen's [52] error taxonomy, all important cognitive mechanisms that lead to human error are explained and extended by further research. The advantages and disadvantages of predictive and cognitive modelling approaches are presented. Subsequently, the present driver behavior model approaches to obtain realistic traffic are explained. Finally, the advantages and application areas of realistic driver behavior models are outlined. The paper shows that the cause of human error and the underlying mechanisms thereof should be considered in order to obtain more transparent and realistic driver behavior models. The driver behavior model DReaM (Driver Reaction Model) is introduced to address the issues resulting from existing cognitive models. In addition, it is shown, which concepts for modelling Rasmussen's error causes and mechanisms currently exist and where further research in needed.