APLICABILIDADE DE MODELOS DE TOMADA DE DECISÃO DE RISCO PARA ENTENDER O COMPORTAMENTO DE MOTRISTAS DURANTE RETOMADAS DE CONTROLE EM AUTOMAÇÃO VEICULAR


 
 
Este trabalho apresenta uma apreciação teórica da aplicabilidade de modelos de tomada de decisão de risco como uma ferramenta para entender o comportamento de motoristas durante retomadas de controle de um veículo automatizado. O artigo se foca na relação entre o conceito de “Out of the Loop” e consciência da situação. Uma discussão metodológica é feita, e suas implicaçóes para o design de produtos é apresentada. Ao fim da discussão, este artigo conclui que o processo de acumulação de evidência em modelos de tomada de decisão possui paralelos fortes com o conceito de retomada de consciência da situação. Dito isto, modelos de acumulação de evidência podem ser tuilizados como ferramentas para entender como motoristas usam a informação para tomar decisões seguras, e esta informação pode ser reforçada no design de interfaces embarcadas. Ao fim do artigo, um modelo conceitual é apresentado como sugestão para aplicação prática da teoria proposta em dados experimentais. 
 
 



INTRODUCTION
Among the human factors-related challenges of implementing vehicle automation, is ensuring safe responses from users during transitions of control. Recent research into this issue forms part of a larger body of research regarding the better design of human-machine interfaces, spanning multiple domains and decades. These challenges highlight an old irony of automation, where the more reliable the automation, the less prepared the human is to react in a time of need (Bainbridge, 1985). This is especially true for higher levels of vehicle automation, which do not require continuous monitoring of the driving task, but still rely on users to resume control, for example, when a system limitation is reached (Level 3. See SAE, 2018 for a complete description of the levels of vehicular automation).
Many recent driving simulator studies, for example, those described by , have identified that drivers in higher levels of vehicle automation (SAE L2+), are removed from the decision-making and control loops of the driving task, placing them "out of the loop" (see  for a recent description of the term). This disengagement from the loops is thought to reduce drivers' capacity to react in dangerous situations, increasing the likelihood of collisions.
Many researchers have tried to understand what constitutes a safe transition of control from automation, investigating what factors influence the success of a transition. For example, Gold et al. (2013) demonstrated that drivers' response to an impending collision, following a request for a transition of control, is dependent on the amount of time given to drivers for this response. These authors report that when drivers were given less time to react, they reacted faster, but more erratically, as shown by the vehicle's lateral and longitudinal accelerations. In contrast, when given more time to respond to an impending collision, drivers reacted more slowly but had a more stable response profile. Zeeb at al. (2015Zeeb at al. ( , 2016 have shown that drivers' take-over time and the quality of this take over (measured as vehicle lateral deviation), is linked to their attention to the road environment during automated driving, with higher levels of distraction to other, non-drivingrelated tasks, leading to a deterioration of take-over quality. However, Louw et al. (2018) suggest that take-over time and vehicle controllability alone are not good predictors of a safe transition of control, but rather the early mitigation of a threat, with earlier transitions of control leading to fewer collisions.
A common limitation of studies attempting to correlate drivers' visual attention with their performance on non-driving-related tasks during automation, is that most investigate the location of drivers' gaze, rather than attempting to understand how visual information, acquired from different sources during automation engagement, affects drivers' resumption of control. While there have been efforts to model the factors that influence drivers' capabilities to take-over control, and how they use the physical and mental resources they need to perform such an action, most have not managed to generate a predictive model, based on gaze patterns during take-overs (Happee et al., 2018). For example, in Victor et al. (2018), while have reported that some drivers, even though looking to the road centre, still failed to avoid crashes during a transition of control (similar to results also reported by . Studies in other domains have considered how visual information sampling affects decision making in humans (see Orquin & Loose, 2013 for a complete literature review of these studies). For instance, Fiedler & Glöckner (2012), identified that gamblers shift their gaze towards the gamble they are willing to make, before their decision, and used this information as a predictor of their choice selection. This paper proposes that the application of decision making theories, and related models, can be used to address some of the gaps in research on user resumption of control from vehicle automation, by providing a quantifiable method of linking the acquisition of specific information from the environment to the probability of a particular response (Orquin & Loose, 2013). Currently, there are only a few studies that highlight the possibility of such a link (c.f. Markkula et al., 2018). In this work, we consider how theoretical models for risky decisionmaking can be used to study drivers' transition of control in automation by observing their visual sampling behaviour during different stages of the take over process.
We begin with outlining the two theoretical bases of this work: decision-making theory, and the human factors of transitions of control. Thereafter, the two theories will be compared, especially regarding their analogous processes of Situation Awareness acquisition and evidence accumulation. Finally, this paper considers how such an approach can generate outputs that may be applied by system designers, to enhance driver performance and create safer systems.

TRANSITIONS OF CONTROL FROM VEHICLE AUTOMATION
This section of the paper aims to define key concepts in the field of human factors of transitions of control, such as the decision-action loop, Situation Awareness, and the issues that are related to this process. With a clear definition of this concept in hand, it will be possible to compare them to the concepts related to the decision-making theory, understanding how they might interact and complement each other.
The term transition of control was described by Louw (2017) as: "the process and period of transferring responsibility of, and control over, some or all aspects of a driving task, between a human driver and an automated driving system." SAE (2018) complement this definition with a taxonomy, by outlining how a driver's responsibility varies across the different levels of automation, and a distinction if they were system-or driver-initiated transitions. The need for such transitions of control is partly based on current system limitations, in terms of the technology's operational design domain (see NHTSA, 2016, for a more descriptive definition of the problem), where vehicles cannot operate in all scenarios, and the human drivers are expected to supervise the automation and resume control, whenever a system limitation is reached. However, the inherent problem with such supervisory roles is diminished driving capabilities associated with the relinquishing of control, which his associated with several challenges when drivers are requested to resume control, especially in time-critical scenarios (Louw, 2017). Some of these issues are discussed below.

The decision-action loop
According to many authors (e.g. Young, 2012), manual driving is a task which requires the driver to always be in the information processing "loop", with regards to their interactions with the surrounding road environment, as well as their ability to control and coordinate vehicle manoeuvres, involving steering, acceleration and braking. Thomas (2001) states that the operation of a vehicle is closely associated with constant feedback and feed-forward cycle of human interaction with the task. Here, humans' decisions and actions affect the situation, and this change is perceived once more by the individuals, who orient and adjust their behaviour accordingly.  further complement this logic for the context of vehicle automation (based on the model purposed by Michon, 1985), by stating that there are two distinct loops in manual driving, which can be affected by ceding control to automation: one for motor-control coordination, and another for the several decision-making processes that need to be performed while driving. They suggest "(…) that "being in the loop" can be understood in terms of (1) the driver's physical control of the vehicle, and (2) monitoring the current driving situation (…)" . It must be noted that both loops continually interact with each other, and drivers must be aware of both their visual-motor coordination (see Wilkie et al., 2008 for a more descriptive definition of the term) and the surrounding environment, to safely maintain control of the task.

Situation Awareness Recovery
Using driving simulator experiments, Louw et al. (2016), supplemented by previous evidence from Damböck et al. (2013), argue that by removing drivers from the decision-making and control loops, vehicle automation reduces drivers' Situation Awareness (SA; Endsley, 1995), which needs to be re-acquired in order to safely resume control and avoid potentially dangerous situations on the road (Damböck et al., 2013). The definition of Situation Awareness used in this research, and defined initially by Endsley (1988), is: "the perception of the elements in the environment within a volume of time and space, the comprehension of their meaning, and the projection of their status in the near future." In short, SA can be divided into three levels (perception; comprehension and prediction), which allow humans to orient their decisions in a particular context and volume of time (Fig. 2).  Figure 2 Endsley's model of SA. This is a synthesis of versions she has given in several sources, notably Endsley (1995) and Endsley et al. (2000), in Wickens (2008).
The loss of Situation Awareness and its relation to being "out of the loop" have been declared by a number of studies on vehicle automation (Carsten et al., 2012;Ohn-Bar & Trivedi, 2016;Morando et al., 2019), some of which have considered how these concepts are affected by drivers' engagement in non-driving-related tasks. It is argued that upon a request to resume control from automation, drivers have to move their visual attention from the NDRT, to focus on other sources of information, related to the driving task, to acquire enough SA to take back control of the vehicle. Gartenberg et al. (2014) refer to this process (which is not only relevant to vehicle automation) as Situation Awareness Recovery or SAR. This is described as a visual scanning process with a considerable number of short fixations in different areas, with a significant lag of resumption in tasks, and a high probability of re-fixation to the same information source, more than once. Examples of such a process was observed in Louw et al. (2019), who reported in their driving simulator experiments that drivers who were engaged in a visual nondriving-related task during automation (assumed to induce an OotL state) had a more scattered gaze pattern after resumption of control from a silent automation failure, compared to those who were required to monitor the road environment during automation.
One of the challenges for the human factors community in addressing this problem is that the process of SAR is accompanied by several barriers, called SA challenges (Endsley, 2006). Endsley & Kriss (1995) named several challenges for the Situation Awareness acquisition, such as attention tunnelling, change blindness, stress on operators' (drivers') working memory, as well as the division of the information required from multiple sources, making it difficult for operators to gather all the information they might need in a reasonable amount of time (e.g. see Parasuraman & Riley, 1998). For driving automation, it has been demonstrated that time pressure, or information overload, might affect the quality of drivers' performance. This is thought to be because drivers' attentional resources are continuously stretched by the high demands of the driving task itself, which is aggravated by automation (Goodrich & Boer, 2003). The dispersion of drivers' gaze also competes between focused attention to the vehicle's heading (due to a visual-motor coordination, Wilkie et al., 2008) and hazard perception routines, which are generally characterised by an increased lateral gaze dispersion (Crundall et al., 1999). Therefore, drivers not only have to acquire information about the situation in the environment, and the current status of the system (an issue also reported by Endsley, 2006), but also have to recover their visual-motor coordination, which is degraded once you relinquish control from the vehicle (Mole et al., 2019). Many empirical studies show that this need to disperse visual attention to different sources affects drivers' performance, increasing risk of crashes (see Russel et al., 2016;Zeeb et al., 2015;Blommer et al., 2016;Merat et al., 2014;Gold et al., 2013;Damböck et al., 2013).

DECISION-MAKING THEORY PRINCIPLES AND MODELS
The definition of decision-making adopted in this work was proposed by Edwards (1954), and is defined as follows: "(…)given two states, A and B, into either one of which an individual may put himself, the individual chooses A in preference to B (or vice versa)". This definition was further developed by Simon (1959), who added organised this process into four main stages: 1) definition of the problem, 2) identification of possible solutions, 3) objective assessment of the value of each solution for the problem, 4) choice of the best solution. As human beings, we are continuously making decisions, based on our internal representation of what we should do in every situation, given certain parameters (stage 3). In a driving task, many actions involve a decision-making process. Some examples include deciding: a comfortable car-following distance (Boer, 1999), what gaps we will accept when changing lanes (Gipps, 1986), how we respond to a potential forward collision (Blommer et al., 2017), and whether to disengage from automation (see Markkula et al., 2018, for more examples).
In the context of this paper, decision-making can be defined as the drivers' choice to take-over control of the vehicle or not, and their take-over modality (how do they take-over). When constructing a model for such decision-making, to account for a good or bad decision, in terms of safety, we have as observable output variables the decision-making time (how long drivers took to decide to take-over), decision choice (how they reacted to the given scenario) and outcome (based on the objectives established for the given situation, were they able to achieve this goal?). Yet, there are several kinds of decision-making theory models, which may account for different aspects of human behaviour, and might be useful for certain situations and not others. Edwards (1954) also divided the decision-making theory models in two main spectrums, which their most recent and developed definitions shall be further explained in the later sections of this paper: the rational and risky decision-making models.

Rational decision-making models
The concept of rational decision-making (see Simon (1979) and March (1978) for a more descriptive definition of the term) is based on a metaphorical "thinking man", as a decisionmaker. According to Simon (1979) and March (1978), a thinking man can be characterized as an individual by two main conditions: 1) as being capable of acquiring and distinguishing all possible relevant information for the decision in hand; and 2) the thinking man is capable of assigning the correct value of a specific choice, based on their established goal in each decision-making scenario. Based on these assumptions, two individuals would always arrive at the same conclusion, when making a rational decision about the same problem. The only Good examples of rational decision-making models can be seen in game theory (Nash, 1950), which posits that all choices made by an individual have a counterpart by a "hostile" opponent (like a chess game). The opponent will focus their actions on maximising their chances of achieving their goal, which is the opposite of the individual's goal. Another example of a rational decision can be seen in the utilitarianism theory, created by Jeremy Bentham and John Stuart Mill in the early 19th century. This theory holds that there are "greater goods" in life, and every moral action can be quantified in terms the outcome of "happiness", and that it is always right to maximise happiness in our choices in life for a "greater good" (for a more complete description of the term, see Mill, 1868). Indeed, rational decision-making processes are utopic in most cases, and their scope for applicability is limited, as everything needs to be quantifiable, such as in mathematical logic problem solutions (for examples, see Bell et al., 1988).

Risky decision-making models
According to decision-making theory, whenever the decision-maker is forced to make a decision without a clear notion of the possible outcomes of their choice, this process is considered to be a risky decision (Edwards, 1954). Models in the risky decision-making theory are based on the assumptions: 1) that not all variables can be accurately, or even wholly, quantified, 2) that humans are not certain about how their actions will affect the environment of the task in hand, and, 3) humans are not aware of are all the variables that they should consider to make their decision. Humans in that situation can estimate, based on their mental models (see Nielsen, 2010 for a description of the term), the probable outcomes for a given task for each possible action that they can perform, and use that information to guide their decisionmaking. In situations where the outcome of an individual's decision is not predictable, they need to account for a level of uncertainty as part of their decision-making process. Uncertainty is defined by Shaw (1983) as the inability of the decision-making to assign the correct value of an option, nor predict the outcomes of their decision to the given environment. This uncertainty concept is a key assumption underlying risky decision-making models and is discussed later in this paper. As humans' mental processing is not directly observable, risky decision-making models can be used to explain human behaviour based on certain assumptions. The most relevant ones are described below: Evidence accumulation models assume that every decision-maker a priori does not have sufficient information about the situation to make a decision and will seek evidence that will influence their decision towards one of the outcomes known to them. Furthermore, every individual has a personal threshold of accumulated evidence that, once reached, causes them to opt for one possible choice, over another (Ratcliff & Smith, 2004). This threshold varied based on a number of factors, including experience, gender, personal attitudes and many others. It must be noted that the rate of evidence, or "drift", is accumulated differently for every person, which is also influenced by a number of factors. In the field of vehicle automation, Markkula et al. (2018) have demonstrated how to apply decision-making models based on evidenceaccumulation to explain, for example, what information drivers use to decide how to resume control from vehicle automation to avoid an incoming forward collision.
Bounded rationality models, first defined by Simon (1972), which holds that humans can make decisions based on the information available to them. These have similar assumptions to rational decision-making models but differ in that they assume that humans are not capable of considering all the relevant information to make a decision. This can be caused by a lack of cognitive resources, time pressure, or simple lack of knowledge about the presence of a particular source of information. Considering this paradigm, bounded rationality models assume that the decision-maker prioritises certain information over others (randomly or selectively). This prioritised information will most likely bias the decision towards a particular choice, depending on the information sampled, and not only on individual preferences. This kind of model is especially relevant for the transition of control in vehicle automation, as it is assumed that drivers in such situations can be overloaded with large volumes of spatially dispersed visual information, and may not be able to process all the information they would need Examples for such overload can be found in Gold et al. (2013) and Blommer et al. (2017), who show that drivers change their decisions about when to resume control from automation, based on the amount of time they have to react before the automated system reaches its limit. Although, it is worth considering that those authors haveonly considered visual information, so other factors might also have affected the observed results.
Satisficing decision-making models assume that the decision-maker will not seek the most optimal solution for his/her problem, but instead will make the first decision where the outcome satisfies their needs or goals in the given situation (Wierzbicki, 1982;Parke et al., 2007). This approach was used in studies by Boer (1999), Boer & Hoedemaeker (1998), and Goodrich & Boer (2003), in different scenarios. For example, Boer (1999) demonstrated that drivers tend to have not one specific "ideal car-following distance", but rather have a satisficing margin, that floats closer or further to the lead vehicle, where the drivers assume to be safe and close enough to be satisfied and refocus in other demands from the car-following task (such as lateral control of the vehicle), instead of actively re-adjust their following distance to a point they would consider to be ideal.
Most concepts in these models are somewhat interchangeable and can be combined in a descriptive or mechanistic analysis. Their relationship with the field of automation will be discussed in the subsequent sections of this work.

RELATIONSHIP BETWEEN HUMAN FACTORS CHALLENGES AND RISKY DECISION-MAKING
Based on the two types of decision-making theory models described above, it is evident that the process of Situation Awareness recovery during the transition control from vehicle automation presents several similarities to the risky decision-making theory, which is discussed in the following sections.  stated that drivers re-enter the cognitive loop of the driving task by acquiring sufficient levels of Situation Awareness. In the same way, Ratcliff & Smith (2004) claim that whenever an individual is presented with an opportunity to make a decision, they will need to accumulate evidence that will support the choice they eventually make. This direct comparison shows similarities in the applicability of both the concept of evidence accumulation and SA for those theories with the same purpose, which is to understand how humans use the information to react to a given environmental condition and achieve their desired goal. Fig. 2 presents a schematic representation of the proposed relationship between the two theories.  Figure 3 Representation of the relationship between SA and decision-making theory As mentioned above, decision-making theory holds that the decision-making process is composed by four steps: 1) define the problem, and understand its characteristics; 2) formulate/generate possible solutions for the given problem; 3) estimation of the value of possible outcomes; 4) selection of the outcome with the highest value for the given problem (see Simon, 1959 for a better description). Endsley (1995) divided the SA into levels, in a way that the individual needs to 1) identify the elements in the environment, 2) comprehend their meaning, and how it shapes the situation in hand, and 3) orient how those elements can be interacted with, in a way that is possible to predict what can be the outcomes of their potential actions. According to Simon (1957) and Edwards (1954), a decision can only be made if there is a clear notion/definition of the value of each solution to the upcoming problem, and that to achieve this, the decision-maker accumulates evidence that assigns the correct value to a particular option, reducing the decision-maker's level of uncertainty (Shaw, 1982). Observing the same phenomenon through the lenses of the SA theory, we can understand that the comprehension of the problem (in the case of this work, a request to transition control) and their possible solutions as level two SA. The process of assigning value, or expected outcome of possible action in order to make the appropriate decision can be directly linked to the level three situation awareness, or projection of future states. In this framework, it can be assumed that the process of moving from level two to level three SA can be directly compared to the process of accumulation of evidence, which is simply the reduction of uncertainty about the outcomes of a possible action to a given scenario.
The arguments presented in the previous section showed that barriers, called SA challenges (Endsley, 2006), impede an individual's ability to acquire all the sufficient levels of SA they need to make an optimal resumption of control from automation (see Parasuraman & Riley (1997) for an example of such phenomenon). Analysing the challenges imposed to an individual to resume control from automation through the lens of decision-making theory, similar problem is reported by Edwards (1954) and Simon (1957) who say that an entirely rational decision is  Blommer et al. (2017) and Gold et al. (2013) showed that drivers have an increased probability to "just brake", instead of both braking and steering, whenever they had limited time to respond to the scenario. The authors noted that the scenario exceeded drivers' abilities to cope with the situation and to perform the ideal action. These two examples can be translated in the risky decision-making theory as satisficing decision-making actions, where even if it was not perfect, it was the best they could do with the information they had, opting to make a simple reaction to the scenario. Based on the arguments presented above, we believe that risky decision-making theory is a suitable candidate to model the process of the take-over of control from vehicle automation. The application of decision-making theory can complement the existing studies on the transition, as it can be used to understand the relationship between the information sampled by drivers and their subsequent behaviour. Practically speaking, this approach complements the current studies in the field by providing robust mathematical models that assign causality between evidence accumulation and decision (see Orquin & Loose, 2013), which are not commonly linked to the situation awareness theory. It is now essential to evaluate how this theory can be applied and implemented to better describe driver behaviour during transitions of control. Sivak (1996) stated that vision is the most important of the five human senses for driving, but yet, it is not suited to dealing with multiple demands at the same time. For this reason, drivers need to prioritise certain visual information over others to perform a transition of control (for more details about this process, see Goodrich & Boer, 2003).

USING DECISION-MAKING MODELS TO ORIENT DRIVERS' DECISION-MAKING
According to Orquin & Loose (2013), visual attention and decision-making are tightly coupled, since a driver's risky decision-making is continuously biased by whether or not they attended to relevant visual information available to them. In their literature review, the authors found a co-causal relationship between visual attendance to information and the occurrence of specific choices, in a discrete decision-making scenario. As part of a meta-analysis, the authors analysed several decision-making tasks that used eye-tracking data as a dependent variable. They concluded that an individual's gaze fixation on certain essential information could predict their upcoming choice in a discrete scenario, suggesting that the selective attention of drivers may bias their decision-making. Such an approach may also be applied to analyse drivers' response capabilities in a take-over scenario, once a take-over reaction is nothing more than a selective response to a particular scenario condition.
The arguments above support the possibility of modelling the relationship between different gaze allocation strategies and the probability of yielding specific responses to the takeover control scenario (based on the studies reported by Orquin & Loose, 2013). This approach would inform system designers about which information should be scanned with higher priority, to yield a higher probability of safe and timely responses to different take-over scenarios. This information could be used to create HMIs that guide drivers towards making decisions that result in safe outcomes. For example, indicating where drivers should focus their attention for a