Bayesian Predictive Model for Electric Level 4 Connected Automated Vehicle Adoption

Abstract

Electric Level 4 connected automated vehicles (CAVs) are now allowed to demonstrate their automation capability in shared mobility robotaxi and microtransit services in geofenced areas in several cities around the world. Private and public sector stake-holders need predictions of their adoption without regulatory constraints for personal mobility and use in shared mobility services. In anticipation of the future presence of CAVs in transportation vehicle fleets, governments are planning necessary regulatory and infrastructure changes. Accompanying this need for forecasts is the acknowledgement that CAV adoption decisions must be made under uncertain states of technology and infrastructure readiness. This paper presents a Bayesian predictive modelling framework for electric Level 4 CAV adoption in the 2030–2035 application context. The inputs to the Bayesian model are obtained from effectiveness estimates of CAV applications that are processed with the Monte Carlo method to account for uncertainties in these estimates. Scenarios of CAV adoption in the 2030–2035 period are analyzed using the Bayesian model, including the quantification of the value of new information obtainable from demonstration studies intended to reduce uncertainties in technology and infrastructure readiness. The results show that in the 2030–2035 application context, the CAVs are likely to be adopted, provided that the trajectory of progress in technology and infrastructure readiness continues, and potential adopters are offered opportunities to learn about Level 4 CAV technological capabilities in a real life service environment. The threshold level of the probability of adoption enhances significantly with high-reliability demonstration results that can reduce uncertainties in adoption decisions. The findings of this research can be used by private and public sector interest groups.

1. Introduction

Automation in driving continues to attract research, development, and demonstration activities around the world. Substantial investments have already been made by the private as well as the public sector, and more are needed to further advance the technology and related infrastructure readiness for adopting electric Level 4 connected and automated vehicles (CAV) without regulatory constraints. The Society of Automotive Engineers (SAE) International has defined the five levels of automation noted below [1].

• Level 1: Driver Assistance
• Level 2: Partial Driving Automation
• Level 3: Conditional Driving Automation
• Level 4: High Driving Automation
• Level 5: Full Driving Automation

Some new-model vehicles with automation Levels 1 to 2 are already on the market. At present, the availability of Level 3 automated vehicles on the market is very scarce, and regulatory approvals are needed for their use. Level 4 CAVs are now allowed to demonstrate their ability to operate in shared mobility ride-hailing robotaxi services in geofenced areas in several cities around the world. Also, demonstrations of slow-speed Level 4 shuttles in microtransit services are underway with similar constraints. In these services, a safety technician is either onboard or monitors operations remotely [2,3]. Public road and highway authorities have initiated the process of accommodating automated vehicles, but so far there is a lack of uniformity in infrastructure changes and regulations.

The adoption of Level 4 CAV is of research interest in this paper. It features “high driving automation; the sustained and operational design domain (ODD)-specific performance by an advanced driving system (ADS) of the entire dynamic driving task (DDT) and DDT fallback without any expectation that a user will need to intervene” [1].

Looking ahead, Level 4 CAVs, have the potential to become mass market transportation solutions for personal mobility and robotaxis for shared mobility [4,5]. Likewise, the growing use of larger Level 4 shared mobility vehicles in microtransit demonstration services is an indicator of their future potential [6,7,8,9].

At an advanced level of development and when fully supported by necessary infrastructure, electric Level 4 CAVs have the potential to offer benefits to private owners, shared mobility commercial service investors (i.e., robotaxi fleet owners), and public agencies as well as private sector investors interested in their use in microtransit services. For this reason, the potential purchasers and deployers of Level 4 CAVs continue to be keen on learning about timing estimates for their implementation [9]. Insights on this complex subject are provided by published literature, but there is a need for a new predictive modelling framework that can address uncertainties in technology and infrastructure readiness in forecasting electric Level 4 CAV adoption [9].

This paper presents research on a predictive model for electric Level 4 CAV adoption for the 2030–2035 market context. Specifically, the research is intended to meet the following objectives:

(1)

Predicting the effectiveness of electric Level 4 CAVs under uncertain states of technology and infrastructure readiness in three applications, namely (a) private mobility, (b) robotaxi shared mobility, and (c) shared mobility microtransit services.

(2)

Addressing uncertainties in effectiveness estimates using the Monte Carlo method and producing two probability-weighted expected effectiveness estimates for each CAV application, one corresponding to a relatively low level of technology and infrastructure readiness, and another corresponding to a higher level of readiness. These are needed as inputs to the Bayesian predictive model.

(3)

Formulation and implementation of the Bayesian model for predicting the probability of CAV adoption in 2030–2035 application scenarios of technology and infrastructure readiness states, including the quantification of the value of new information obtainable from demonstration studies intended to reduce uncertainties in the readiness states.

(4)

Obtaining insights from the predictive model results on the conditions under which CAVs are likely to be adopted in the 2030–2035 period.

Although a Level 4 CAV is of interest in the industry as a freight vehicle, this application is beyond the scope of research reported in this paper.

Following this introduction, the methodological framework and its components are described. The functions of technology in their respective applications are explained. Next, the multi-criteria effectiveness method is applied to Level 4 CAV services, and the Monte Carlo method is used to account for uncertainties in effectiveness estimates. These results become inputs to the Bayesian predictive model. The theoretical foundation of the Bayesian model is explained, and applications are illustrated. The discussion section covers methods and results. Finally, conclusions are presented.

2. Methodological Framework

Attempts have been made in the past to study factors for defining timing estimates for CAV implementation in various service contexts. These studies were reported by public agencies or societies [9,10], independent researchers [11,12,13], consultants [14,15,16,17,18], and financial and related institutions [19,20,21,22,23,24,25]. A review of past studies suggests observations of interest to this research:

• Several interest groups are keen on learning about CAV adoption forecasts. These include vehicle manufacturers and marketers, public sector infrastructure owners and operators, private sector transportation companies, financial institutions and investors, consulting firms, researchers, and consumer groups.
• Factors that could be used to forecast decisions on CAV adoption for various applications include technological capabilities, infrastructure for supporting CAV use, government regulations, differences between CAV and non-CAV travel characteristics, trends in general consumer acceptance of automation in driving, technology costs, and investor sentiments.
• With no substantive CAV market in place today, a methodological framework for forecasting must rely entirely on informed subjective estimates of CAV application effectiveness.
• Treating uncertainty in all of the inputs to a predictive model is necessary. Likewise, all parts of the Bayesian model support decision-making under uncertainty.

Given the current interest in forecasting many facets of CAV applications, the modelling framework shown in Figure 1 is defined and used in this research. Here, a brief overview of the framework and its components is noted, and theoretical and methodological details are provided in the following sections of the paper.

The first two boxes in Figure 1, labelled as “inputs”, define necessary information on technology and service context, characterize uncertain states of technology and infrastructure readiness, define effectiveness criteria, and outline use of the Monte Carlo simulation method to obtain the probability-weighted expected effectiveness estimates that are required by the Bayesian predictive model. The third box shows parts of the Bayesian model. Taken together, these provide answers to questions on the conditions under which CAV adoption will become likely and the role of new information obtainable from demonstrations in reducing uncertainty in CAV adoption for three applications, namely private mobility, shared mobility robotaxi, and shared mobility microtransit service.

The methodological framework development was guided by the following requirements:

• Working with effectiveness criteria that can be quantified largely in non-monetary terms (e.g., user satisfaction, technology malfunction).
• Due to uncertainties in technology and infrastructure readiness and user acceptance of CAVs (i.e., for personal mobility, as a shared mobility vehicle), the method should be able to work with a range of criteria achievement levels.
• Use of a widely utilized method to account for uncertainties in criteria achievement levels and to produce probability-weighted expected effectiveness outputs required as inputs to the Bayesian predictive model.
• Need for a predictive model with the following necessary capabilities:

  (a)

  Applying probabilities to uncertain states of technology and infrastructure readiness.

  (b)

  Enabling a role for new information on uncertain variables, obtainable from demonstrations.

  (c)

  Updating probabilities of uncertain states of technology and infrastructure readiness using the new information.

  (d)

  Producing answers as to the value of new information in reducing uncertainties and enhancing the basis for CAV deployment decision for services defined above.

3. Overview of Level 4 CAV Technology and Infrastructure

Level 4 CAVs’ high automated driving ability is noted in the introductory section of this paper [1]. Communication technologies and equipped traffic control methods will be needed for supporting connected vehicle applications. Although an automated vehicle can function using traffic signals that are not digitized, digital infrastructure can improve efficiency and safety. An automated vehicle can operate using existing infrastructure with the use of cameras and sensors to interpret a control setting. Digitally enhanced systems such as those used for vehicle-to-everything (V2X) can be used for real-time information exchange.

Transportation agencies are advancing V2X communications, which enable direct communications between devices. Dedicated short-range communications (DSRC), cellular (4G/5G), and potentially satellite networks have relevant roles to play in CAV operation [26,27,28,29].

For technology and supporting infrastructure to be considered to be at an advanced stage of development, significant CAV technology advances as well as extensive supporting technology implementation in the road traffic network will be needed. Some forecasters go beyond these requirements by suggesting the integration of mobile devices carried by pedestrians and bicyclists into the V2X ecosystem [26].

Public and private sector investments in automation in driving research, development, and demonstration (RDandD) have not yet advanced the state of technology to the extent that deployment of Level 4 CAV can be planned with certainty of only positive effects and avoidance of negative impacts.

For the foreseeable future, it is necessary for technology regulators and investment decision-makers to work with uncertain states of technology and infrastructure readiness. In the following sections of the paper, in the 2030–2035 service context, forecasts of positive as well as negative attributes of the Level 4 CAV technology are described for two uncertain states: relatively low readiness of technology and infrastructure and relatively high readiness of technology and infrastructure.

4. 2030–2035 Service Context

4.1. Private Mobility Vehicle

This vehicle will be owned and maintained by the owner. The batteries can be charged at several locations (i.e., at the residence of the owner, at a work location, at a commercial fast charging station, etc.) [30]. These vehicles will be used for all travel purposes in the city, on intercity routes, and on rural roads. Availability of communication service will enable the use of automation as well as connected features of the technology. The presence of a human driver in the driver’s seat strengthens the safety attribute of the Level 4 CAV private mobility vehicle.

4.2. Robotaxi

Following regulatory approval, fleets of electric Level 4 CAVs can be deployed in ride-hailing robotaxi services without geofence limitations. Passengers can request a CAV using an app. The robotaxi system, also known as a shared automated vehicle (SAV) system, consists of vehicles, stations for parking, battery charging, and, among other functions, devices to monitor and control the CAV fleet. As a commercial business, these systems are expected to become economically feasible [31].

Even in their current state of development, shared automated vehicle applications are planned around the world [32]. Robotaxis operate in well-defined geofenced areas in several cities around the world without a safety driver in the vehicle (but in a control center). The technology has advanced to the extent that it makes it unnecessary to have a human safety driver in the vehicle. Instead, a remote-control center is used for surveillance and emergency correction of the driving functions. These specially designed centers are connected with robotaxis, using a combination of wireless communication technologies, for real-time support and guidance in situations that exceed their automation capabilities.

The communication technologies used include cellular networks (4G/5G), V2X communication, and potentially satellite communication for locations with limited cellular coverage. Human operators in the control center monitor service operations and resolve potential issues. The communication system enables them to provide navigation assistance, re-route the vehicle, or manage unexpected conditions. The knowledge that the remote human operator can assist the CAV contributes to user trust in the system.

4.3. Microtransit Service

Microtransit, as a shared mobility system, is experiencing rapid growth. It is defined as “Privately or publicly operated, technology-enabled transit services that typically use multi passenger/pooled shuttles or vans to provide on-demand or fixed schedule services with either dynamic or fixed routing” [33,34]. Many public transit service gaps are addressed by microtransit, including the first/last mile connectivity that enables travelers to complete the origin-to-destination trip [35].

In 2030–2035, an appropriately sized electric Level 4 CAV will be able to provide a cost-effective on-demand microtransit service without a safety driver in the vehicle. Avoiding the driver’s cost greatly enhances the cost-effectiveness of the service. Research shows that the driver cost for an accessible battery electric non-automated minibus with 19 seats accounts for 43% of vehicle cost [35]. Although the Level 4 CAV-based microtransit system is monitored by a control center, its cost does not adversely affect the feasibility of the service [33]. For information on the communication technologies that enable the CAV fleet to connect with the control center and the role of the human operator in the control center, please see the above robotaxi section.

Given the potential of Level 4 CAV-based microtransit systems for cost-effective services, the public transit industry and urban governments are looking forward to their mass-produced availability for microtransit services to fill the well-known gaps in urban public transit networks [31,36,37].

5. Level 4 CAV Technology and Infrastructure Attributes

5.1. Positive Attributes (2030–2035 Service Context)

Following necessary public sector regulatory approvals, Level 4 CAV deployments can be planned. As noted earlier, in this research, the following are of interest: personal mobility, robotaxi service, and CAV-based microtransit service. Details of these applications are presented in Section 4 of the paper. Here, the attributes of technology and the associated infrastructure are described. These, in essence, serve as effectiveness criteria for CAV deployments and are key variables that influence trust in and acceptance of CAVs.

The following four positive effects of electric CAV applications are commonly described in the literature:

• Human factor-related collisions avoided (safety benefit).
• User satisfaction (as owner of the personal passenger vehicle, user of the robotaxi service, or user of the CAV-based microtransit service).
• Socio-economic benefits (other than safety benefits).
• Environmental benefits.

The main rationale for favorable public policies regarding automation in driving is to reduce traffic accidents caused by human errors [27,28,38]. The CAVs can potentially avoid collisions due to a combination of advanced sensors, algorithms, and communication with vehicles and infrastructure [39]. According to a UK source, human errors account for 88% of all road collisions, making CAVs a potentially transformative safety technology [40].

User satisfaction with Level 4 CAVs is a multifaceted attribute. Surveys of consumers show that they like the availability of automation features [14], but their acceptance is unlikely without testing the technological capabilities using a demonstration vehicle. Potential owners can benefit from greater levels of safety (covered above). Convenience features add to user satisfaction, including the ease of operation for parking, merging, and other maneuvers. Numerous publications have covered the subjects of user acceptance and trust models [41,42,43,44]. These studies found the main variables to be the same as the ones defined here as positive attributes and avoiding the negative attributes described in the following section.

Traveler acceptance of the robotaxi service remotely monitored by a control center is subject to uncertainty. Likewise, accepting a microtransit shuttle vehicle monitored by a remote-control center may be viewed differently than travelling in a slow-speed demonstration shuttle with a technician in the vehicle. Therefore, user/traveler acceptance of Level 4 CAVs cannot be assumed with certainty.

More than sixty factors of expectation, experience, and acceptance of automation in public transportation were identified by review articles [42,43,44]. Frequently cited desirable service factors include seat availability, comfort, being on time, schedules, fares, road safety, onboard safety and security, and cyber security. Traveller personal factors are also known to influence their acceptance (e.g., socio-demographics, travel habits, and personality).

Interviews of demonstration shuttle vehicle users carried out before and after travel, in general, did not show major concerns. Models developed suggest acceptance of automated shuttles in the future, subject to absence of technical issues and a clearer explanation of legal responsibilities in the case of an accident [42,43,44].

The CAVs offer recognized socio-economic benefits, including increased safety (covered above). Specifically, some socio-economic benefits noted by survey respondents are improved mobility for those who are unable to drive due to age, disability, or other reasons, enhanced accessibility for underserved populations, and affordable shared mobility [28,45].

Electric CAVs offer environmental benefits when used for personal mobility and in shared mobility services [32]. To maximize environmental benefits, large-scale market penetration of electric CAVs as well as availability of battery charging infrastructure are necessary. Towards this end, increased public and private sector efforts are needed to secure public and consumer trust [46].

5.2. Negative Attributes (2030–2035 Service Context)

The following four negative effects of electric Level 4 CAV applications are commonly described in the literature:

• Technology unreliability.
• Effect on other road users.
• Hacking and data security.
• Cost differential (i.e., extra cost of automation).

Although in the 2030–2035 service context, the Level 4 CAV technology and supporting infrastructure are expected to be more advanced than now, there may still be a need for further development (i.e., these technologies may not be considered “mature”). Surveys of potential adopters and technology experts suggest that at the present state of development, technology unreliability is a concern. Further advances will be necessary before a Level 4 CAV will be ready to avoid collisions in edge cases and therefore serve the market without regulatory conditions. In this context, potential users consider occupants’ safety and legal liability in case of an accident as issues that should be addressed [41,47].

The need to improve the reliability of sensors in Level 4 CAVs has been noted in the literature. Therefore, continued R&D efforts are needed so that these become fully ready for widespread deployments in all environments. At their present state of development, these CAVs operate safely in specific and well-defined conditions in geofenced areas without a safety driver onboard; however, for the future increased scope of operations, technology advances and supporting infrastructure readiness are necessary [32,48].

Crash data for CAV Levels 3–5 show that as automated vehicle usage increases, crashes increase as well, and technology malfunction may be one of the causes [49]. Survey results suggest that public concern about technology unreliability increases with the automation level [46]. Studies by the American Automobile Association (AAA) Foundation for Traffic Safety show that the public as well as technology experts recognize the importance of the safety attribute of CAVs and that they are concerned about technology malfunction [50,51]. Other surveys found similar safety concerns [32].

The effect of CAV operation on other users of the road needs designer attention. Given that multiple accidents have occurred between automated vehicles and pedestrians, this concern is not without evidence [32]. Recognizing the importance of minimizing the adverse effects of CAV operation on vulnerable road users, R&D efforts continue to address challenges in their interaction. Specifically, efforts are reported to improve technology (e.g., algorithms, training, datasets) for enhanced accuracy in all types of interactions [52].

Cybersecurity and data privacy will continue to be challenges in future years because Level 4 CAVs feature a large attack surface which makes them vulnerable to cyberattacks. Although the connectivity attribute of the CAV system offers safety and convenience benefits, without safeguards, it becomes a vulnerability that can be exploited by hackers [53,54,55].

The cost difference between a CAV and a vehicle without connected and automated technologies in the 2030–2035 application period is another attribute of interest in this research. The consumer survey results suggest concerns with the purchase and on-going maintenance costs of Level 4 CAVs [50]. A forecasting study suggests that automation will add 20% to the cost of a vehicle [56]. Although the cost differential between electric CAVs and conventional technology vehicles will improve in favor of CAVs, it is not likely to be zero. The price difference is expected to considerably decrease over time due to the following reasons: technology advancements, increased competition, and mass production. In the case of shared mobility services, all costs (i.e., purchase, operations, maintenance, and management costs) are expected to drop due to higher vehicle utilization, increased ridership, wider operating areas, and service optimization [32].

6. Uncertain States and Decision-Making Under Uncertainty

The decision to adopt a Level 4 CAV will be made by a decision-maker or decision-makers under uncertainty. It is assumed that the decision to adopt the CAV for individual mobility will be made by the owner. The decision to adopt the CAV for robotaxi services will be a corporate decision. Likewise, the decision to adopt the CAV shuttle for microtransit services will be made by the fleet procurement manager on behalf of the organization. The variables for modelling the decision-making problem are shown in

Table 1. Variables for modelling the decision-making problem.

Use Cases (2030–2035 Market Context)	Decisions	Decision-Makers	Uncertain States	Impacts (Effects)
Personal passenger vehicle use	A1: Purchase CAV A2: Do not purchase CAV	Individual consumer	S1: Uncertain. Low state of technology and infrastructure readiness S2: Uncertain, but higher state of technology and infrastructure readiness	Multi-attribute utility of A-S combinations.
Robotaxi service	A1: Invest in CAV fleet, infrastructure, operation and management system. A2: Do not.	Corporate decision	S1: Uncertain. Low state of technology and infrastructure readiness. Developing market potential. S2: Uncertain, but higher state of technology and infrastructure readiness. Promising market potential.	Multi-attribute utility of A-S combinations.
CAV-based microtransit Service	A1: Invest in CAV fleet, infrastructure, operation and management system. A2: Do not.	Public transit management decision. Corporate decision for private sector service.	S1: Uncertain. Low state of technology and infrastructure readiness. Developing market potential. S2: Uncertain, but higher state of technology and infrastructure readiness. Promising market potential	Multi-attribute utility of A-S combinations.

. It is understood that government regulations will allow the sale and use of CAVs for the licensed service. Although regulations may permit demonstration services in limited conditions as is the case now, these permits do not imply commitment to allowing full-scale deployments without further technology and supporting infrastructure improvements.

Many technology and infrastructure attributes can be assessed to infer the state of readiness for Level 4 CAV adoption. Key attributes are the safety and reliability of technology (including occupant safety and security, safety of other road users), a favorable record of cybersecurity and safeguarding data, successful human–machine interfaces as required for Level 4 CAV operation (including verification that in the case of shared mobility services, the remote control staff can indeed resolve issues), consumer acceptance of the cost of automation in driving, resolution of legal responsibility in case of an accident, and a favorable record of operating permits granted by regulatory authorities without geofencing and other constraints.

The state of technology and infrastructure readiness in the 2030–2035 CAV implementation period cannot be predicted with certainty. There are many key actors who play RD&D and permitting roles according to their objective function. These include vehicle producers and providers, infrastructure owners and operators, private transportation companies, financial institutions and investors, consulting and strategy firms, researchers, public officials and politicians, and consumers [9].

Technological forecasting studies such as the one reported in this paper must work with relevant influencing factors for which objective data may not be available. Therefore, reliance is placed on subjective estimates. In this research, despite efforts in sourcing objective data/estimates, it is necessary to rely on informed subjective information regarding the following variables: trajectory of technological capabilities, government regulations, shifting travel behavior, perception of benefits obtainable from CAVs as compared with non-automated vehicles, general consumer trends, technology costs, and investor sentiment [9].

The impacts/consequences/effects of a combination of an adoption decision A and a state of technology and infrastructure readiness S (i.e., A&S) are quantified in relative value (utility) metrics [57]. As explained in the following section of the paper, there are well-established theoretical and practical reasons for using utility metrics in this study of CAV adoption for the applications noted in Table 1.

7. Quantifying the Effectiveness of CAV Application in Meeting Owner/User Criteria

7.1. Utility (Relative Value) Theory

In Section 5.1 and Section 5.2, eight attributes of Level 4 CAV application are discussed. Since these characterize the effects of adopting the CAV (e.g., collisions avoided, user satisfaction, technology malfunction incidents), these serve as effectiveness criteria. Although some criteria can be quantified in their respective units (e.g., accidents avoided, number of technology malefaction incidents), others will require relative value units (e.g., user satisfaction, many and diverse socio-economic benefits). Also, market values do not exist for most effects. For these reasons, these effects cannot be quantified in monetary units or in units of any other criterion. Well-recognized methods are available for researching this type of socio-technical problem. Following the study of manuals and application examples, the multi-criteria analysis method was adopted in association with utility metric to quantify the effectiveness of Level 4 CAV applications for personal mobility, shared robotaxi services, and shared mobility microtransit services [38,57].

The effectiveness of a CAV application can be expressed as

$$e={\sum }_{i=1}^q p_i({Cr}_i)·e({Cr}_i)$$

(1)

where

• e is the effectiveness of the Level 4 CAV application.
• Cr_i is the achievement level, by the CAV, of a criterion i, i = 1,2, …, q.
• e(Cr_i) is the utility of achieving Cr_i (e.g., collisions avoided).
• p_i(Cr_i) is the probability that Cr_i will be achieved by the CAV application (e.g., robotaxi).

The methodology permits the consideration of differential effects (e.g., safety of various vulnerable users of the road). In such a case, the hth level of criterion g, Cr_gh, which can occur due to CAV operation, can be expressed as

Cr_gh = the hth level of criterion g (e.g., safety), weighted for all affected groups

= k₁Cr_g¹ + k₂Cr_g² + k₃Cr_g³ + ….. + k_sCr^v_g

(2)

where

Cr_g^v = the level of achievement of criterion g for group v (e.g., safety of group v—pedestrians—etc.)

k_v = a weight, reflecting the importance of the impact group v with respect to criterion Cr_g, and can be determined from the societal (community’s) preference expressed as ranks such that ${\sum }_{v}k$_v = 1.0.

The scale for measuring positive criterion achievement is 0 to 100, and for negative criterion, 0 to −100 is used. The criterion achievement values (called effectiveness values) are based on potential technology improvements (i.e., past or current observations cannot be used). It is a very challenging but essential task to assign values for various levels of criteria achievement (e.g., utils for avoiding collisions). Commonly available statistics, user surveys, or opinions of experts are used as guides for this purpose.

The quantification of criterion achievement levels follows the axioms of the utility theory [57]. As noted above, the utility value is a relative measure of the degree to which each criterion is achieved by a CAV application. Since positive and negative scales are used in this research, the utility theory permits the addition of values measured on positive and negative value scales.

Due to the lack of objective information on Level 4 CAV technology and supporting infrastructure in the 2030–2035 application context, single values of criterion achievement cannot be assigned with certainty. Therefore, the following methodological steps are applied to account for uncertainty:

• The effectiveness values are estimated as ranges, the higher the degree of uncertainty, the wider are the estimates.
• The Monte Carlo method, described in the following section, is applied to calculate the probability-weighted expected value of criterion achievement.
• Next, the expected value of criterion achievement level obtained from Monte Carlo simulation can be transformed, if warranted.

The process of transforming the expected criterion achievement level obtained from the Montecarlo method into utility units (utils) is illustrated in Figure 2. Three types of utility functions are shown that can be used for scale transformation purposes: a non-linear function that exhibits diminishing marginal value, a linear transformation function with an intercept, and another linear function without an intercept.

For the non-linear case,

e(Cr_i) = H[Cr_i (Cr_i measured on the original value scale)]^z for z < 1 and I = 1,2, …, q

(3)

where

• e(Cr_i) = the utility measure, in transformed units, for criterion Cr_i.
• H and z are constants.
• For the case of linear transformation with an intercept,

e(Cr_i) = m_iCr_i + b_i for i = 1,2, …, q

(4)

where

• m_i = the slope of the transformation curve for Cr_i
• b_i = the vertical axis intercept (if applicable, a threshold step can be applied).

The origin of the scale can be set in any desirable way. For the example evaluations presented in this paper, 0 ≤ e(Cr_i) ≤ 100.

Although in explaining the theory, and as illustrated in Figure 2, three utility functions are noted, due to lack of experience with electric Level 4 CAV applications, only linear transformation without intercept is used. However, for future use of the methodology, it is useful to note that the utility functions can be derived from stated preference or revealed preference surveys of consumers/interest groups. If sufficient data become available to estimate the non-linear utility function, it may exhibit the property of diminishing marginal utility.

The expected effectiveness of an electric Level 4 CAV application j in 2030–2035 is obtained as

e_j = ∑ _{i = 1 to q} [w_i·(eCr_i)]

(5)

where

• e_j = the expected effectiveness of CAV application j, weighted for all impact/interest groups
• w_i = the weight assigned to criterion i

The expected effectiveness method is applied to three electric Level 4 CAV applications: a personal mobility vehicle, a shared mobility robotaxi service, and a shared mobility microtransit service. Details of input information, intermediate computations, and results are reported in Section 7.3.1, Section 7.3.2, Section 7.3.3 and Section 7.3.4.

7.2. Montecarlo Method to Treat Uncertainties

The Monte Carlo method uses random numbers to sample probability distribution functions for computing expected value and standard deviation. In this research, instead of using the middle values of the range of uncertain effectiveness estimates, these are analyzed using the scientific Monte Carlo simulation method. The use of this theoretically sound and well-recognized scientific method was suggested by the US DOT for the IntelliDrive benefit–cost analysis [58]. An introduction to simulation and Monte Carlo methods can be read in Rubinstein and Kroese [59], and Evans et al. describe characteristics of probability distribution functions and their suitability for various applications [60]. The Monte Carlo method is suitable for use in this research due to its ability to work with a range of values of uncertain variables (i.e., minimum and maximum values), application of probabilities sampled from specified distributions, and if applicable, treating peaking of values.

A rectangular (also called uniform) probability distribution function and a triangular probability distribution function are well suited for the analysis of stochastic effectiveness variables and are therefore used in this research (Figure 3 and Figure 4). The Monte Carlo method draws samples from their cumulative probability distribution functions using random numbers.

The continuous uniform probability distribution function is characterized by two values: the minimum value α and the maximum value β. This distribution is used to represent a higher uncertainty regarding the phenomenon under study by treating all values of the random variable as equally probable. In scientific terms, this probability distribution is the maximum entropy probability function for a stochastic variable denoted as X [59].

The statistical characteristics of the continuous uniform probability distribution function are noted below.

Probability density function: P(X) = 1/(β − α) for X ϵ |α,β| and 0 otherwise

(6)

Mean: ½(α + β)

(7)

Median: ½(α + β)

(8)

Mode: any value in (α,β)

(9)

The continuous triangular probability distribution function is defined by the minimum value α, the maximum value β, and the peak (i.e., the mode or most likely) value λ. This probability distribution function has been widely applied in real-life problem-solving conditions under uncertainty due to its ability to treat the minimum and maximum values and the most frequent outcome (i.e., the mode). These values can be assigned without knowing the difficult-to-obtain mean and the standard deviation of the values of the variable of interest. By assigning definite lower and upper limits, the analyst can bypass the assumed extreme values. Further, this function enables the treatment of skewed probability distributions [59].

The main statistics of the triangular probability density function are noted below.

P(X) = 2(X − α)/[(β − α)(λ − α)]

(10)

for α ≤ X ≤ λ and

2(β − X)/[(β − α)(β − λ)] for λ ≤ X ≤ β

Also, P(X) = 0 for X < α and X > β

Where λϵ|α,β| is the mode

(11)

The mean is 1/3[(α + β + λ)]

(12)

7.3. Computation of Criteria-Weighted Effectiveness

7.3.1. Effectiveness Criteria Weights

In Section 7.1, the theoretical foundation of the effectiveness method is described. Here, components of Equation (5) are developed. The criteria weights suggested for use in this research are shown in

Table 2. Effectiveness criteria and criteria weights.

Criteria	Criteria Weight	Normalized Weight
Cr1 Human factors-related collisions avoided	10	0.29
Cr2 User satisfaction	3	0.09
Cr3 Socio-economic benefits	2	0.06
Cr4 Environment	2	0.06
Cr5 Tech unreliability	4	0.12
Cr6 Effect on road users	6	0.18
Cr7 Hacking, data security	4	0.12
Cr8 Cost differential	3	0.09
	34	~1.00

. Safety and security criteria are assigned higher importance than other criteria. These weights reflect surveys of potential users of electric Level 4 CAVs. Examples of notable surveys that were carried out are those by the Foundation for Traffic Safety of the American Automobile Association (AAA) [46,51] and McKinsey & Company [32,61]. The raw weights on a 0–10 scale are normalized for use in Equation (5).

7.3.2. Private Vehicle Effectiveness

In accordance with the methodology, the criteria effectiveness values are assigned on a 0 to 100 scale for criteria Cr1 to Cr4 and on a 0 to −100 scale for criteria Cr5 to Cr8. The values shown in

Table 3. Effectiveness under S1: Level 4 CAV as a private vehicle (utils).

Criteria	S1 Effectiveness (−100 to +100 Scale)	S1 Expected Effectiveness & St. Dev. Based on Triangular Probability Distribution	S1 Expected Effectiveness & St. Dev. Based on Uniform Probability Distribution	S1 Weighted Effectiveness (Based on Uniform Probability Distribution Results and Normalized Weights)
Cr1 Collisions avoided	25 to 40	32.7 and 4.5	33.1 and 4.6	9.6
Cr2 User satisfaction	30 to 45	37.5 and 3.4	37.6 and 4.7	3.4
Cr3 Socio-economic benefits	15 to 30	22.5 and 3.4	22.6 and 4.2	1.4
Cr4 Environment	20 to 40	30.0 and 5.5	30.0 and 7.1	1.8
Cr5 Tech unreliability	25 to −50	−50.5 and 6.2	−39.2and 7.4	−4.7
Cr6 Effect on road users	−35 to −65	−44.3 and 7.0	−50.4 and 9.0	−9.1
Cr7 Hacking, data security	−25 to −50	−38.3 and 5.7	−39.2and 7.4	−4.7
Cr8 Cost differential	−60 to −90	−76.0 and 5.4	−75.4 and 9.2	−6.8 −9.1

reflect conditions under the low state of technology and infrastructure readiness scenario S1 for a Level 4 CAV. This is the pessimistic view of technology projections for the 2030–2035 period.

Table 4. Effectiveness under S2: a Level 4 CAV as a private vehicle (utils).

Criteria	S2 Effectiveness (−100 to +100 Scale)	S2 Expected Effectiveness & St. Deviation Based on Triangular Probability Distribution	S2 Expected Effectiveness & St. Deviation Based Uniform Probability Distribution	S2 Weighted Effectiveness (Based on Uniform Probability Distribution Results and Normalized Weights)
Cr1 Collisions avoided	40 to 70	55.1 and 6.3	55.2 and 9.8	16.0
Cr2 User satisfaction	30 to 45	37.8 and 3.3	37.8 and 4.6	3.4
Cr3 Socio-economic benefits	20 to 40	30.0 and 5.5	30.0 and 7.1	1.8
Cr4 Environment	0 to −25	30.0 and 5.5	30.0 and 7.1	1.8
Cr5 Tech unreliability	0 to −30	−12.3 and 5.1	−12.6 and 7.4	−2.5
Cr6 Effect on road users	0 to −20	−15.2 and 6.9	−15.6 and 9.8	−2.8
Cr7 Hacking, data security	0 to −20	−10.7 and 4.6	−10.8 and 6.9	−1.3
Cr8 Cost differential	−10 to −40	−26.3 and 6.8	−25.1and 8.1	−2.3 14.1

presents the inputs and results for the scenario of a higher state of technology and infrastructure readiness S2 (i.e., the optimistic view). It is suggested that criteria achievement estimates under S1 and S2 should be viewed jointly for an appreciation of the uncertainties shaping these projections for a new technology and its application.

The literature cited in previous sections provided sufficient information for suggesting the relative effectiveness values. As previously noted, the effectiveness values (+ve and −ve) are estimated as a range of values due to estimation under uncertainty. The rationale for assigning the criteria attainment values presented in Table 3 and Table 4 is noted next.

• Cr1 Human factor-related collisions avoided: The values range from 25 to 70 utils. Survey results clearly suggest that potential adopters as well as technology experts are calling for increased safety in vehicle design and improved road infrastructure. Also, the need for “more regulations” is noted [46,61]. According to the Victoria Transport Policy Institute, “many technical problems must be solved before autonomous vehicles can operate reliably in all normal conditions” [12]. The low end of 25 utils collision avoidance effectiveness is plausible for Level 4 CAVs under the S1 scenario that reflects operation of an under-developed vehicle without the required infrastructure support. On the other hand, the high effectiveness level is achievable under S2, which characterizes a state free of technical problems [46,61].
• Cr2 User satisfaction: The range is from 30 to 45 utils under both S1 and S2. A CAV that may be allowed to operate in the 2030–2035 service environment is likely to have a moderate level of capability that users require (e.g., automated longitudinal and lateral movements, parking assistance) [46,61].
• Cr3 Socio-economic benefits: The overall range of 15 to 40 utils reflects a general lack of experience in obtaining socio-economic benefits from the new technology. Some authors include driving-related savings in this category of CAV effectiveness. In relative terms, the CAV under S2 has higher potential to generate benefits than under S1 due to favorable technology and infrastructure developments [12,45,61].
• Cr4 Environment-related effects: The range of 20 to 40 utils is based on survey information that showed that about 12% of EV owners are well set in terms of chargers and another about 28% are getting close to having enough chargers needed [61]. Under S2, the environment-related benefit is higher than S1 due to higher level of infrastructure readiness.
• Cr5 Technology unreliability: Potential adopters and experts have concerns regarding technical issues. About 56 percent of survey respondents were concerned about technology malfunction [61]. The ranges defined (i.e., −25 to −50 utils under S1 and 0 to −25 under S2) reflect opinions expressed by numerous authors, survey information, and the belief that CAV technology is likely to improve over time [9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,61].
• Cr6 Effect on road users: Although there are no survey data on the percentage of experts concerned about the impact of Level 4 CAVs on other road users, research indicates that their safety is a valid concern. [12]. The effectiveness range under S1 is from −35 to −65 utils, but it improves to 0 to −30 utils under S2.
• Cr7 Hacking and data security: 46% of survey respondents are concerned about vehicle hacking and 55% show data privacy as an issue with Level 4 CAV technology and operations [46]. These statistics are reflected in the −25 to −50 utils range under S1. However, technology and operational improvements provide sufficient basis for the 0 to −20 utils range under S2 [12,46].
• Cr8 Cost differential: 68% of respondents are extremely or very concerned about the purchase price of Level 4 CAVs [46]. Additionally, there are concerns with maintenance costs [9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25]. The effectiveness range of −60–90 utils under S1 reflects much concern. However, due to dropping costs, the range improves to −10 to −40 utils under S2.

In columns 3 and 4 of Table 3 and Table 4, the expected effectiveness values and St. deviations results of the Monte Carlo method are shown. The results obtained with the use of the uniform probability distribution function generally reflect higher uncertainty as compared to the triangular probability distribution function. In the last column of Table 3 and Table 4, the weighted effectiveness values are based on uniform probability distribution results and normalized weights shown in Table 2. In accordance with the axioms of the utility theory, the utility values can be added.

In Table 3, −9.1 utils is the answer obtained for Equation (5). The negative sign indicates that if the low state of technology and infrastructure readiness S1 becomes true, the use of a Level 4 CAV as a private vehicle is not favorable.

As noted above, Table 4 presents the inputs and results for Level 4 CAVs for the scenario of a higher state of technology and infrastructure readiness (i.e., S2). That is, if S2 becomes true, the effectiveness values and the final weighted effectiveness answer of 14.1 utils for Equation (5) shown in

Table 5. Effectiveness under S1: a Level 4 CAV as a ride-hailing robotaxi (utils).

Criteria	S1 Effectiveness (−100 to +100 Scale)	S1 Expected Effectiveness & St. Deviation Based on Triangular Probability Distribution	S1 Expected Effectiveness & St. Deviation Based Uniform Probability Distribution	S1 Weighted Effectiveness (Based on Uniform Probability Distribution Results and Normalized Weights)
Cr1 Collisions avoided	25 to 40	32.7 and 4.5	33.1and 4.6	9.6
Cr2 User satisfaction	25 to 45	37.5 and 4.5	33.3 and 4.6	3.0
Cr3 Socio-economic benefits	15 to 30	22.6 and 3.1	22.8 and 4.2	1.4
Cr4 Environment	20 to 40	30.0 and 5.5	30.0 and 7.1	1.8
Cr5 Tech unreliability	−25 to −50	−38.3 and 5.7	−39.2 and 7.4	−4.7
Cr6 Effect on road users	−30 to −60	−44.3 and 7.0	−43.3 and 9.6	−7.8
Cr7 Hacking, data security	−25 to −50	−38.3 and 5.7	−39.2 and 7.4	−4.7
Cr8 Cost differential	−50 to −80	−65.3 and 7.5	−66.7 and 9.8	−6.0 −7.4

are likely to occur. Therefore, under S2, CAV use as a private vehicle is favorable.

7.3.3. Ride-Hailing Robotaxi Effectiveness

Table 5 and

Table 6. Effectiveness under S2: a Level 4 CAV as a ride hailing robotaxi (utils).

Criteria	S2 Effectiveness (−100 to +100 Scale)	S2 Expected Effectiveness & St. Deviation Based on Triangular Probability Distribution	S2 Expected Effectiveness & St. Deviation Based Uniform Probability Distribution	S2 Weighted Effectiveness (Based on Uniform Probability Distribution Results and Normalized Weights)
Cr1 Collisions avoided	40 to 70	55.1 and 6.3	55.2 and 9.8	16.0
Cr2 User satisfaction	30 to 45	37.2 and 3.1	37.7 and 4.1	3.4
Cr3 Socio-economic benefits	20 to 40	30.0 and 5.5	30.0 and 7.1	1.8
Cr4 Environment	20 to 40	30.0 and 5.5	30.0 and 7.1	1.8
Cr5 Tech unreliability	0 to −25	−12.2 and 5.1	−12.7 and 7.1	−1.5
Cr6 Effect on road users	0 to −30	−15.2 and 6.9	−15.6 and 9.8	−2.8
Cr7 Hacking, data security	0 to −30	−15.0 and 6.8	−15.0 and 8.6	−1.8
Cr8 Cost differential	−10 to −30	−20.9 and 4.0	−20.4 and 5.9	−1.8 15.1

provide inputs and results for the Level 4 CAV used for a ride-hailing robotaxi service. Since this service is based on Level 4 CAVs, the survey information obtained from references [46,61] and the rationale for assigning effectiveness values described for the private vehicle are applicable to the robotaxi case. However, there are minor differences in effectiveness values due to application of the vehicle as a robotaxi.

The private vehicle owner is likely to experience a somewhat higher degree of satisfaction than a robotaxi user. The robotaxis are monitored by professional staff, and, in comparison, the human driver in the private mobility vehicle may not always be fully alert in case a corrective action is needed that the automated system may not be able to handle. Another difference is regarding the cost of automation. Due to fleet discounts the robotaxis are likely to be better off than a private vehicle. Finally, under both the S1 and S2 states, the ride hailing robotaxi is slightly more effective than the private mobility vehicle.

7.3.4. Microtransit Vehicle Effectiveness

The effectiveness values for the Level 4 CAV used for microtransit services under S1 and S2 states are presented in

Table 7. Effectiveness under S1: a Level 4 CAV used as a microtransit vehicle (utils).

Criteria	S1 Effectiveness (−100 to +100 Scale)	S1 Expected Effectiveness & St. Dev. Based on Triangular Probability Distribution	S1 Expected Effectiveness & St. Dev. Based on Uniform Probability Distribution	S1 Weighted Effectiveness (Based on Uniform Probability Distribution Results and Normalized Weights)
Cr1 Collisions avoided	25 to 40	32.7 and 4.5	33.1 and 4.6	9.6
Cr2 User satisfaction	25 to 40	32.7 and 4.5	33.3 and 4.6	3.0
Cr3 Socio-economic benefits	20 to 40	30.0 and 5.5	30.0 and 7.1	1.8
Cr4 Environment	20 to 40	30.0 and 5.5	30.0 and 7.1	1.8
Cr5 Tech unreliability	−25 to −50	−38.3 and 5.7	−39.2 and 7.4	−4.7
Cr6 Effect on road users	−30 to −60	−44.3 and 7.0	−43.3 and 9.6	−7.8
Cr7 Hacking, data security	−25 to −50	−38.3 and 5.7	−39.2 and 7.4	−4.7
Cr8 Cost differential	−50 to −80	−65.3 and 7.5	−66.7 and 9.8	−6.0 −7.0

and

Table 8. Effectiveness Under S2: Level 4 CAV as a microtransit vehicle (utils).

Criteria	S2 Effectiveness (−100 to +100 Scale)	S2 Expected Effectiveness & St. Deviation Based on Triangular Probability Distribution	S2 Expected Effectiveness & St. Deviation Based Uniform Probability Distribution	S2 Weighted Effectiveness (Based on Uniform Probability Distribution Results and Normalized Weights)
Cr1 Collisions avoided	40 to 70	55.1 and 6.3	55.2 and 9.8	16.0
Cr2 User satisfaction	25 to 40	32.7 and 4.5	33.3 and 4.6	3.0
Cr3 Socio-economic benefits	20 to 40	30.0 and 5.5	30.0 and 7.1	1.8
Cr4 Environment	20 to 40	30.0 and 5.5	30.0 and 7.1	1.8
Cr5 Tech unreliability	0 to −17.5	−8.6 and 4.5	−8.3 and 6.4	−1.0
Cr6 Effect on road users	0 to −30	−15.2 and 6.9	−15.6 and 9.8	−2.8
Cr7 Hacking, data security	0 to −20	−10.7 and 4.6	−10.8 and 6.9	−1.3
Cr8 Cost differential	0 to −30	−14.9 and 5.8	−14.4 and 9.4	−1.3 16.2

, respectively. Since this service is based on Level 4 CAV technology, the survey information sourced from references [46,61] and the rationale for assigning effectiveness values described for the private car and robotaxi are applicable to the microtransit case. However, there are minor differences in effectiveness values due to application of the vehicle to microtransit services.

Due to the high-convenience nature of service, user satisfaction is higher for robotaxis than for microtransit. Microtransit has higher effectiveness scores than robotaxis for hacking and automation cost. The differences arise due to closer monitoring of microtransit CAV operation and management than for robotaxis, a higher fleet discount, and lower operation and management costs. On the balance, the weighed effectiveness score is higher for the CAV being used in microtransit services than for the CAV used as a robotaxi.

8. Bayesian Model: Theory

The technological forecaster needs methods that can identify high-payoff paths under uncertainty. The Bayesian model is widely accepted due to its capacity to work with uncertain states of nature and estimates of gains that are expected to result from action–state combinations. In this research, the uncertain states of technology and infrastructure readiness are S1 and S2 in the 2030–2035 application context, and potential actions are adoption decisions on CAV applications. The weighted effectiveness results (i.e., effectiveness of CAV adoption under S1 and under S2) presented in Table 3, Table 4, Table 5, Table 6, Table 7 and Table 8 become inputs to this method.

Table 9. Variables and processes applied for preparation of inputs for the Bayesian model.

Variables and Processes (1)	Effectiveness Criteria and Weights (2)	Effectiveness Criteria Values (Ranges) (Utils) (3)	Montecarlo Method (Triangular Prob. Distribution) (Utils) (4)	Monte Carlo Method (Uniform Prob. Distribution) (Utils) (5)	Weighted Expected Effectiveness (Utils) (2) × (5)
Processes and results.	Cr1 to Cr8. Raw and normalized weights.	Effectiveness ranges defined.	Computation of probability-weighted effectiveness.	Computation of probability-weighted effectiveness.	Multiplication of (2) × (5) Sum of utils.
Location of results in Table 2, Table 3, Table 4, Table 5, Table 6, Table 7 and Table 8.	Table 2	Table 3, Table 4, Table 5, Table 6, Table 7 and Table 8 column 2	Table 3, Table 4, Table 5, Table 6, Table 7 and Table 8 column 3	Table 3, Table 4, Table 5, Table 6, Table 7 and Table 8 column 4	Table 3, Table 4, Table 5, Table 6, Table 7 and Table 8 column 5

presents a summary of processes used in preparing the weighted effectiveness value inputs to the Bayesian model.

Within the field of statistical methods for decision-making under uncertainty, the Bayesian analysis offers the required capability to model technological forecasting problems. It allows the forecaster to use prior probabilities for the occurrence of uncertain states of nature and offers the ability to update probabilities using new information obtainable from learning methods (i.e., demonstration studies). In the Bayesian method, as explained below, probabilities have important roles in risk analysis.

Considerable efforts by stakeholders from research institutes, industry, and government are underway to provide the latest information to the public on CAVs’ capabilities as well as their limitations. Technical forums provide information and examine the impact of information on consumer understanding of automation in driving. User surveys carried out before and after the use of a demonstration service are useful to gauge change in potential adopter views and confidence.

Given the importance that potential adopters attach to the “need to test it personally”, demonstration of technology is assumed for all analyses. At the time of the demonstration activity, questions on technology, infrastructure, and regulations can be answered. As explained later, the Bayesian model can quantify the change in perceived gain of CAV adoption due to the availability of new information obtainable from learning about technology and infrastructure readiness during demonstrations and related question and answer sessions.

Selected example applications of the Bayesian decision theory to model forecasting problems can be read in References [62,63,64]. These papers show that logical answers were obtained that could not be obtained with other known methods. If a forecast is to be made or a highly desirable course of action is to be identified under uncertainty and it is possible to learn from new information to modify probabilities of the uncertain future conditions, the Bayesian theory is best suited to analyze the problem.

8.1. Prior Analysis

For modelling the CAV adoption decision under uncertainty, the decision maker needs to be known. Table 1 provides this information, and it is repeated here for ease of reference. In the case of Level 4 CAV ownership, the decision maker is the individual consumer. For the robotaxi fleet, the decision maker is the corporation. In the case of a microtransit vehicle fleet, the decision-maker is the public transit agency. On the other hand, in the case of a privately owned system, the investor is the decision maker.

To add to the previous descriptions, variables of the CAV adoption decision model are defined below:

• Alternatives: A1 adopt CAV; A2 do not adopt CAV.
• The condition under which the adoption decision will be made is characterized by the following uncertain states of technology and infrastructure readiness: S1 low state of technology and infrastructure readiness; S2 higher state of readiness.
• The impact (consequence) of each A&S combination is quantified by the expected weighted effectiveness values presented Table 3, Table 4, Table 5, Table 6, Table 7 and Table 8. In Bayesian analyses, these are termed as Gain G of CAV adoption. Depending upon the specific A&S combination, the G can be negative or positive.
• The uncertain states are assigned probabilities of occurrence. These probabilities are called prior probabilities. The term prior is used because the decision maker has the option to ask for a demonstration study, and the new information can be used for revising prior probabilities into posterior probabilities.

In the prior analysis, the prior probabilities are applied to gain G for the applicable A&S combination to calculate probability-weighted expected gain. For the adoption of a selected CAV (i.e., private automobile or robotaxi or microtransit vehicle),

$$G\text{*}(Ai)={\sum }_{j=1 }^2[Gij(Ai)\times P_j]$$

(13)

where

G*(Ai) = the expected gain for alternative $A_i$; $i$= 1 or 2. For example, for the private car,

A1 represents the decision to adopt and A2 is used the decision not to adopt.

$Gij$(Ai) = the gain of alternative $A_i$, under uncertain technology and infrastructure readiness state Sj, j = 1,2

Pj = probability of state of nature Sj (i.e., technology and infrastructure readiness state).

Figure 5 shows the prior analysis part of the Bayesian model.

8.2. Posterior Analysis

The demanding technological forecasting task requires knowledge extraction from available sources as well as the option to initiate new information acquisition activity for use in converting prior probabilities into posterior probabilities. In the decision theory terminology, these initiatives are called learning experiments. In this research the term learning activity (L) is used. At the time of initiating an L about uncertain states of technology and infrastructure readiness, the outcomes are unknown, and therefore, probabilities are applied to their occurrence. Given two unknown states S1 and S2, the results obtainable for L are r1 (that corresponds to S1) and r2 (that corresponds to S2). The option of not initiating a learning activity is L0, and therefore, r0 is the outcome.

The reliability of resulting information about the occurrence of uncertain states (i.e., S1 and S2) is another variable that is needed for revising the probabilities. The revised probabilities are termed posterior probabilities, and their application is known as posterior analysis. In statistical terms, the conditional probability P(r|S,L) enables conversion of prior probabilities to posterior probabilities of technology and infrastructure readiness states.

Equation (13) defined in the prior analysis section enables the analyst to compute the expected gain Gij of an alternative Ai using prior probabilities Pj of uncertain states Sj. The resulting expected gain G is the answer obtained from the prior analysis part of the Bayesian method. Since the prior probability-based analysis does not include a role for new information intended for revising probabilities, the posterior analysis is the next logical step.

Raiffa and Schlaifer [65] defined the theoretical basis of posterior analysis, pre-posterior analysis and the value of new information obtained from a learning activity L. Figure 6 shows the sequence of decision-making under uncertainty in the posterior analysis case.

The G(L,r,A,S) represents gains for all L,r,A,S combinations.

The Bayesian analysis requires the following probability distribution functions:

• Using the prior analysis as a base, the prior probability P′(S) for each S is required before observing the outcome r of the new information acquisition activity L.
• A conditional measure P(r|S,L) is to be assigned, which represents the probability that the result r will be observed if the learning activity L is carried out, and S is the true state. That is, the analyst should define the reliability of the information outcome r of L in predicting the true state S.
• The marginal measure P(r|e) is computed as shown next:

P(r|L) = ΣP′(S) P(r|S,L)

(14)

• The posterior probability P″(S|r,L) can now be calculated using the Bayes Theorem:

P″(S|r,L) = [P(r|S,L)P′(S)]/(P(r|L)

(15)

• This equation reflects the Bayesian philosophy that an L can be characterized by a conditional probability P(r|S,L) (a reliability indicator), used for computing posterior probabilities. The Bayes Theorem (Equation (15)) defines the relationship between the prior and posterior probabilities.

The decision trees for comparing the prior and posterior analyses are shown in Figure 5 and Figure 6. In the posterior analysis, the decision variables are L and A, and random variables are r and S. It is solved by moving from right to left. The value of a sequence of actions is represented by the gain G (L,r,A,S). The prior branch is solved using Equation (13). The equations for solving the posterior branch and a comparison of the results of both branches are presented in the following section. To compare L with L0, decision trees shown in Figure 5 and Figure 6 are analyzed. A comparison of the results of posterior and prior branches leads to the quantification of the value of the new information (i.e., the value of pre-posterior information).

8.3. Pre-Posterior Analysis

The pre-posterior analysis is intended to quantify the increase in gain obtainable from the learning activity L. This materializes due to risk reduction in the choice of the Level 4 CAV adoption. The following are the pre-posterior analysis steps: The starting point is the prior probability P′(S). For the L, the conditional probabilities P(r|S,L) are defined. The marginal probabilities P(r|L) for L are computed using Equation (14). For the null learning option L0, the marginal probability is P(r₀|L0) = 1.0). The posterior probability P″(S|r, L) is computed for each combination of S and r (Equation (15)). For each combination of L,r,A,S, its gain is found: G (L,r,A,S).

The expected gain for each alternative A in the posterior branch is as follows:

G*(A,r,L) = ∑_SP″(S|r,L) G(L,r,A,S)

(16)

However, for the prior branch, where no new information is acquired,

G*(A,r₀,L0) = ∑_s P′ (S) G(L₀,r₀,A,S)

(17)

For each (L,r) combination, the optimal alternative is determined, and its associated gain is noted:

G*(r, L) = Max_A G*(A,r,L)

(18)

For the information acquisition activity L, the expected gain can be computed as follows:

G*(L) = ∑ _r P(r|L) G*(r, L)

(19)

8.4. Value of New Information

The Bayesian decision theory enables the quantification of how much G can be improved (by reducing uncertainty) with new information before obtaining it. It is a unique decision aid that enables the identification of conditions under which to initiate the new information acquisition activity L. The computational steps are noted next and illustrated in Figure 7.

(1)

From the posterior branch, for L and each r, find Max_AG*(A,r,L). Call it Ar.

(2)

For L0 in the prior branch, find MaxG*(A). Call it A′.

(3)

For each r, find (Ar-A′)

This is Vt(L,r), the terminal value for the (L,r) combination

(20)

The subscript t represents terminal values.

(4)

The expected value of new information is computed as follows:

Vt*(L) = ∑_r[P(r|L)(Ar − A′)] = ∑_rP(r|L)V_t(L,r)

(21)

The superscript * represents expected (i.e., probability-weighted) value.

Examples of Vt*(L) application are presented in the following sections. For brevity, these are shown as Vt*.

9. Application of the Bayesian Model

The methodological framework illustrated in Figure 1 shows the process for preparing inputs to the Bayesian model and the steps for model implementation. Details of the Bayesian model are presented in Section 8.1, Section 8.2, Section 8.3 and Section 8.4. Within the Bayesian model, prior probabilities, conditional probabilities, and the gain information are the inputs. The outputs are the value of new information, the preferred adoption option, and the corresponding expected gain value. Figure 8 illustrates the Bayesian model computational steps.

In populating the gains information, the results of the weighted effectiveness computations for the A&S combinations presented in Table 3, Table 4, Table 5, Table 6, Table 7 and Table 8 are used. For example, for a Level 4 CAV used for private mobility, G(A1,S1) = −9.1 utils is used, and the contrary decision is assigned zero gain (i.e., G(A2,S1) = 0 utils). Under S2, G(A1,S2) = 14.1 utils and G(A2,S2) = 0 utils.

An important consideration in assigning zero gain for the A2 (i.e., do not adopt) decision is that no positive or negative effects will result from not adopting the CAV for personal mobility. This philosophical argument is also applied to robotaxi and microtransit cases. The counter argument—that under a low technology and infrastructure readiness state S1, a positive gain (i.e., G(A2,S1) = 9.2 utils)) will occur as the opportunity value if the Level 4 CAV is not adopted—is difficult to justify and therefore is not pursued.

The Bayesian model is used to analyze scenarios for CAV adoption in the 2030–2035 period to define conditions under which CAV adoption will be likely. Also, it is of interest to define conditions under which the value of new information is positive (i.e., when Vt* is > 0). A positive Vt* plays a role in reducing uncertainties and improves expected gain for the adopted Level 4 CAV.

An illustration of the effect of new information (i.e., r2 with a conditional probability of P(r2|S2,L) = 0.7) on changes from prior to posterior probabilities can be observed in Figure 9. For example, starting from a rather low prior probability of P′(S2) = 0.4, in association with the high reliability of favorable new information (i.e., r2), the posterior probability rises to 0.609. This observation is useful in reviewing the results of analyses presented below.

For each Level 4 CAV application under study, the Bayesian model was run using the gain values described earlier. Two types of scenarios were analyzed. First, the prior probabilities were changed while keeping the conditional probability P(r2|S2,L) constant at a high level of 0.7. As noted previously, the conditional probability reflects the reliability of the new information acquisition activity L. Eleven scenarios were run with different prior probabilities of S1 and S2. The P′(S2) was changed from 0.05 to 0.9. Table 9,

Table 10. Bayesian prediction of private auto CAV adoption (A1 adopt, A2 do not adopt).

Prior Prob. P′(S1) and P′(S2)	Conditional Prob. P(r1|S1) and P(r2|S1)	Selected Conditional Prob. P(r1|S2) and P(r2|S2)	Selected Posterior Prob. P″(S1|r2,L) and P″(S2|r2,L)	Gain (Utils) G(A1,S1) and G(A2,S1)	Gain (Utils) G(A1,S2) and G(A2,S2)	Value of New Information Vt*(Utils)	Choice of A and E(A) (Utils)	Interpretation of Results
0.95 and 0.05 to 0.80 and 0.20	0.7 and 0.30	0.3 and 0.70	0.891 and 0.109 To 0.632 and 0.368	−9.1 and 0	14.1 and 0	0	A2 and 0	CAV will not be adopted. New information will not change the choice.
0.70 and 0.30	0.7 and 0.30	0.3 and 0.70	0.500 and 0.500	−9.1 and 0	14.1 and 0	1.05	A2 and 0	CAV will not be adopted. But new information can reduce risk.
0.60 and 0.40	0.7 and 0.30	0.3 and 0.70	0.391 and 0.609	−9.1 and 0	14.1 and 0	2.13	A1 and 0.18	CAV will be adopted, subject to new information for risk reduction.
0.50 and 0.50	0.7 and 0.30	0.3 and 0.70	0.300 and 0.700	−9.1 and 0	14.1 and 0	1.07	A1 and 2.50	CAV will be adopted, subject to new information for risk reduction.
0.40 and 0.60	0.7 and 0.30	0.3 and 0.70	0.222 and 0.778	−9.1 and 0	14.1 and 0	0.01	A1 and 4.82	CAV will be adopted, subject to new information for risk reduction.
0.30 and 0.70	0.7 and 0.30	0.3 and 0.70	0.155 and 0.845	−9.1 and 0	14.1 and 0	0	A1 and 7.14	CAV will be adopted. No need for new information.
0.20 and 0.80	0.7 and 0.30	0.3 and 0.70	0.097 and 0.903	−9.1 and 0	14.1 and 0	0	A1 and 9.46	CAV will be adopted. No need for new information.
0.15 and 0.85	0.7 and 0.30	0.3 and 0.70	0.070 and 0.930	−9.1 and 0	14.1 and 0	0	A1 and 10.62	CAV will be adopted. No need for new information.
0.10 and 0.90	0.7 and 0.30	0.3 and 0.70	0.045 and 0.955	-.9.1 and 0	14.1 and 0	0	A1 and 11.78	CAV will be adopted. No need for new information.

and

Table 11. Bayesian prediction of robotaxi CAV adoption (A1 adopt, A2 do not adopt).

Prior Prob. P’(S1) and P’(S2)	Conditional Prob. P(r1|S1) and P(r2|S1)	Selected Conditional Prob. P(r1|S2) and P(r2|S2)	Selected Posterior Prob. P″(S1|r2,L) and P″(S2|r2,L)	Gain (Utils) G(A1,S1) and G(A2,S1)	Gain (Utils) G(A1,S2) and G(A2,S2)	Value of New Information Vt*(Utils)	Choice of A and E(A) (Utils)	Interpretation of Results
0.95 and 0.05 to 0.85 and 0.15	0.7 and 0.30	0.3 and 0.70	0.891 and 0.109 to 0.708 and 0.292	−7.4 and 0	15.1 and 0	0 to 0	A2 and 0	CAV will not be adopted. New information will not change the choice.
0.80 and 0.20	0.70 and 0.30	0.3 and 0.70	0.632 and 0.368	−7.4 and 0	15.1 and 0	0.338	A2 and 0	CAV will not be adopted. But new information can reduce risk.
0.70 and 0.30	0.7 and 0.30	0.3 and 0.70	0.500 and 0.500	−7.4 and 0	15.1 and 0	1.617	A2 and 0	CAV will not be adopted. But new information can reduce risk.
0.60 and 0.40	0.7 and 0.30	0.3 and 0.70	0.391 and 0.609	−7.4 and 0	15.1 and 0	1.296	A1 and 2.896	CAV will be adopted, subject to new information for risk reduction.
0.50 and 0.50	0.7 and 0.30	0.3 and 0.70	0.300 and 0.700	−7.4 and 0	15.1 and 0	0.325	A1 and 3.85	CAV will be adopted, subject to new information for risk reduction.
0.40 and 0.60	0.7 and 0.30	0.3 and 0.70	0.222 and 0.778	−7.4 and 0	15.1 and 0	0	A1 and 6.1	CAV will be adopted. No need for new information.
0.30 and 0.70	0.7 and 0.30	0.3 and 0.70	0.155 and 0.845	−7.4 and 0	15.1 and 0	0	A1 and 8.35	CAV will be adopted. No need for new information.
0.20 and 0.80	0.7 and 0.30	0.3 and 0.70	0.097 and 0.903	−7.4 and 0	15.1 and 0	0	A1 and 10.6	CAV will be adopted. No need for new information.
0.15 and 0.85	0.7 and 0.30	0.3 and 0.70	0.070 and 0.930	−7.4 and 0	15.1 and 0	0	A1 and 11.725	CAV will be adopted. No need for new information.
0.10 and 0.90	0.7 and 0.30	0.3 and 0.70	0.045 and 0.955	−7.4 and 0	15.1 and 0	0	A1 and 12.85	CAV will be adopted. No need for new information.

present results. In the second type of scenario, the effect of changes in conditional probability P(r2|S2,L) were studied while the prior probability was constant. The results are presented in a later section of the paper.

9.1. Private Automobile

The results of eleven scenarios of Bayesian model prediction of CAV adoption for private mobility are presented in Table 10. The first column shows prior probabilities. For S2, these were varied from 0.05 to 0.90. The second and third columns show conditional probabilities. These were held constant to distill the effect of changes in prior and posterior probabilities. Columns 5 and 6 show gain values for A&S combinations (i.e., G(A,S)). Column 7 presents the value of new information Vt*, and column 8 shows the preferred adoption option and the corresponding expected value E(A). In the final column, results are interpreted.

The value of new information Vt* is positive from P″(S2|r2,L) 0.5 to 0.778. Within this range of probabilities for technology and infrastructure readiness in the S2 state, new information plays a role in reducing uncertainties and therefore improves the expected gain of adoption of A1. CAV adoption (i.e., A1) becomes the choice starting from P″(S2|r2,L) = 0.607, and its expected gain continues to rise from here to the last scenario.

9.2. Robotaxi

The Bayesian model results for the robotaxi CAV are presented in Table 11 using the Table 10 format. The results differ somewhat from those for the private mobility CAV. The Vt* becomes positive sooner, starting with posterior probability P″(S2|r2,L) and becomes zero at P″(S2|r2,L) = 0.778 and follow-up scenarios. This result implies that there is a role for new information to reduce uncertainties earlier than for private automobiles, which helps CAV adoption. The CAV adoption option is favorable starting from P″(S2|r2,L) = 0.609, and the expected gain continues to rise with increased probability of its adoption.

9.3. Microtransit

The Bayesian model results for the Level 4 CAV used for microtransit services are presented in

Table 12. Bayesian prediction of microtransit CAV adoption (A1 adopt, A2 do not adopt).

Prior Prob. P’(S1) and P’(S2)	Conditional Prob. P(r1|S1) and P(r2|S1)	Conditional Prob. P(r1|S2) and P(r2|S2)	Selected Posterior Prob. P”(S1|r2,L) and P”(S2|r2,L)	Gain (Utils) G(A1,S1) and G(A2,S1)	Gain (Utils) G(A1,S2) and G(A2,S2)	Value of New Information Vt*(Utils)	Choice of A and E(A) (Utils)	Interpretation of Results
0.95 and 0.05 to 0.85 and 0.15	0.7 and 0.30	0.3 and 0.70	0.891 and 0.109 to 0.708 and 0.292	−7.0 and 0	16.2 and 0	0 to 0	A2 and 0	CAV will not be adopted. New information will not change the choice.
0.80 and 0.20	0.70 and 0.30	0.3 and 0.70	0.632 and 0.368	−7.0 and 0	16.2 and 0	0.588	A2 and 0	CAV will not be adopted. But new information can reduce risk.
0.70 and 0.30	0.7 and 0.30	0.3 and 0.70	0.500 and 0.500	−7.0 and 0	16.2 and 0	1.932	A2 and 0	CAV will not be adopted. But new information can reduce risk.
0.60 and 0.40	0.7 and 0.30	0.3 and 0.70	0.391 and 0.609	−7.0 and 0	16.2 and 0	0.996	A1 and 2.28	CAV will be adopted, subject to new information for risk reduction.
0.50 and 0.50	0.7 and 0.30	0.3 and 0.70	0.300 and 0.700	−7.0 and 0	16.2 and 0	0.02	A1 and 4.60	CAV will be adopted, subject to new information for risk reduction.
0.40 and 0.60	0.7 and 0.30	0.3 and 0.70	0.222 and 0.778	−7.0 and 0	16.2 and 0	0	A1 and 6.92	CAV will be adopted. No need for new information.
0.30 and 0.70	0.7 and 0.30	0.3 and 0.70	0.155 and 0.845	−7.0 and 0	16.2 and 0	0	A1 and 9.24	CAV will be adopted. No need for new information.
0.20 and 0.80	0.7 and 0.30	0.3 and 0.70	0.097 and 0.903	−7.0 and 0	16.2 and 0	0	A1 and 11.56	CAV will be adopted. No need for new information.
0.15 and 0.85	0.7 and 0.30	0.3 and 0.70	0.070 and 0.930	−7.0 and 0	16.2 and 0	0	A1 and 12.72	CAV will be adopted. No need for new information.
0.10 and 0.90	0.7 and 0.30	0.3 and 0.70	0.045 and 0.955	−7.0 and 0	16.2 and 0	0	A1 and 13.88	CAV will be adopted. No need for new information.

using the formats of Table 10 and Table 11. The results differ somewhat from those for the private automobile and robotaxi CAVs. The value of new information becomes positive at a posterior probability of 0.368 and becomes zero from a posterior probability of 0.778 to the last scenario. In scenarios where the value of new information is positive, it contributes to risk reduction. Starting from a posterior probability of 0.609, based on favorable new information, CAV adoption in microtransit services becomes the preferred alternative. The expected gain continues to increase from a posterior probability of 0.609 to the end of scenarios at posterior probability of 0.955.

9.4. Effect of Posterior Probabilities on Expected Gain of CAV Adoption

As noted above, a Level 4 CAV will be adopted for all applications starting from posterior probability P”(S2|r2,L) of 0.609 and remains the preferred option in all remaining scenarios (Figure 10). The combined expected gain of adopting CAV plus the value of new information increases with increasing posterior probability. In relative terms, the microtransit has the highest gain value, followed by the robotaxi and the private mobility automobile has the lowest gain. These results are logical due to the ranks observed from the weighted effectiveness values (Table 3, Table 4, Table 5, Table 6, Table 7 and Table 8).

9.5. Role of New Information Obtained from Learning Activity

In technological forecasting, in general, there is always a need for updating information on uncertain variables. This is the case with the Bayesian predictive model. The decision maker can use new information on sources of uncertainty and use the results for more informed decision-making. In this study, the expected value of new information increases the expected gain of adopting a CAV for each intended application.

Figure 11 illustrates the role of the new information shown in Table 9, Table 10 and Table 11. The expected Vt* for each CAV application was computed for scenarios with altered prior probabilities, but conditional probabilities were held constant (i.e., the reliability of the learning activity was held constant). Since prior probabilities were altered, the computed posterior probabilities changed. In the figure, the Vt* are shown for various posterior probabilities P″(S2|r2,L).

Several observations can be drawn from Figure 11. For the very low probabilities of technology and infrastructure readiness state S2, there is no need for new information. The same is the case for high probabilities of S2. In these scenarios there is no role for new information to reduce uncertainty and improve expected gain. However, for several scenarios shown in the figure, new information can reduce uncertainty and therefore add value to the expected gain.

For the scenarios illustrated, all CAV applications benefit from risk reduction (i.e., these have a positive value of new information). In relative terms, the private mobility CAV benefits the most from risk reduction and the shared mobility applications have generally comparable Vt* values.

9.6. Effect of Reliability of New Information

On the assumption that in 2030–2035, S2 (i.e., the high state of technology and infrastructure readiness) will become true, the corresponding outcome r2 of an information acquisition activity L should be able to indicate the same, based on the following indicators:

• Technology developers, infrastructure providers, and regulators are on track for improving the safety and reliability of technology and services. Also, detailed information is available on Level 4 CAV technology capabilities and limitations.
• Plans are underway to improve cybersecurity and safeguarding data.
• The potential owner of the Level 4 CAV for private mobility can personally test its automation functions, including the call for human intervention, if needed.
• The potential users of shared mobility Level 4 CAVs intended for robotaxi and microtransit services are offered the opportunity to verify that the remote-control staff can indeed resolve issues.
• By 2030–2035, the automation costs will improve in favor of Level 4 CAV ownership, maintenance, and operations.
• Detailed information is available regarding legal responsibilities in case of an accident.

If the learning activity L has high potential to provide comprehensive and reliable information regarding the future state S2 in terms of the above indicators, it can be considered to have a high conditional probability P(r2|S2,L). As noted earlier, this probability definition reads as follows: “given that S2 will become true in the future, what is the probability that the outcome r2 of L will indicate the same”.

At the time of initiating new information acquisition activity, it is uncertain what the outcome will be. That is, obtaining result r1 or r2 will be uncertain, and therefore, probabilities should be assigned to each. As noted previously, r1 corresponds to state S1 (low technology and infrastructure readiness) and r2 corresponds to S2 (relatively higher readiness). If L has high reliability, the analyst can assign a high conditional probability P(r2|S2,L). If L has limited scope in terms of providing information on the above noted indicators, a low conditional probability can be assigned.

To test the effect of the reliability of answers obtainable from the learning activity L, 

Table 13. Effect of conditional probabilities on Vt*.

P(r2|S2,L)	P′(S1) and P′(S2)	P″(S2|r2,L)	Private Auto Vt*	Robotaxi Vt*	Microtransit Vt*
0.5	0.6 and 0.4	0.40	0	0	0
0.6	0.6 and 0.4	0.50	1.02	0.248	0
0.7	0.6 and 0.4	0.609	2.13	1.296	1
0.8	0.6 and 0.4	0.727	3.24	2.344	2.06
0.9	0.6 and 0.4	0.857	4.35	3.392	3.13

presents five scenarios. Column 1 shows that the conditional probability (that characterizes the reliability of L) varies from 0.5 to 0.9. Contents of column 2 show that prior probabilities are held constant so that change in value of new information Vt* due to change in conditional probability can be observed. Column 3 shows calculated posterior probabilities of S2 that correspond to the r2 outcome of L. The last three columns show the computed values of new information for the three applications of the Level 4 CAV.

The results show that at a conditional probability of 0.5, Vt* is zero for all applications (i.e., there is no role for new information obtained from a low reliability L in reducing risk). This is also the case for a conditional probability of 0.6, but only for microtransit. Another observation is that the value of new information increases with increased conditional probability. In relative terms, the private mobility CAV benefits the most from risk reduction and microtransit has the lowest need for risk reduction obtainable in the form of new information. These results are illustrated in Figure 12.

10. Discussions

The research reported here has benefitted from numerous technological forecasting and behavioral user trust and acceptance studies. These provided the basis for quantifying the effectiveness of an electric Level 4 CAV as a personal mobility vehicle and in shared mobility robotaxi and microtransit services. The analytical and statistical natures of methods included in the methodological framework have enabled the use of informed subjective estimates of future values of stochastic variables. This approach is in line with observations of other researchers who reported factors for use in studies on timing estimates for automated vehicle implementation.

The results of the multicriteria effectiveness and Monte Carlo methods are well suited as inputs to the Bayesian predictive model used for identifying conditions favorable to Level 4 CAV adoption for personal use and in shared mobility applications in the 2030–2035 period. The two types of scenarios tested, namely changes in the posterior probabilities of technology and infrastructure readiness states and the conditional probabilities (that characterize the reliability of new information in reducing uncertainties), provided sufficient information to answer questions on CAV adoption decisions under uncertainty. The results are logical.

The expected gain of adopting the Level 4 CAV (in all application cases) increases with increased prior probability P′(S2), corresponding posterior probability P″(S2|r2,L), and improved reliability of new information (i.e., increase in conditional probability P(r|S,L)). For improved robustness of Level 4 CAV adoption results, higher conditional probabilities are desirable.

Based on the assumption of a reliable new information acquisition activity (i.e., real life demonstration with conditional probability P(r|S,L = 0.7), the posterior probability P″(S2|r2,L) of approximately 0.61 is the adoption threshold for all CAV applications (i.e., for personal mobility and shared mobility). These results are logical, given that the expected effectiveness of Level 4 CAVs under S2 is positive and much higher than under S1 (see Table 3, Table 4, Table 5, Table 6, Table 7 and Table 8). Also, these results imply favorable conditions for Level 4 CAV adoption.

In relative terms, the microtransit application has the highest gain, followed by robotaxi, and the lowest expected gain is for a private mobility CAV. Due to the nature of variables included in the expected effectiveness/gain formula, the gain can be regarded as a proxy for user and societal benefits.

In general, the need for new information to reduce uncertainties, revealed by Vt*, is higher for personal mobility Level 4 CAV adoption than for applications in shared mobility services. The robotaxi service is likely to require a higher level of trust from its potential adopters than the microtransit service.

11. Conclusions

Automation in driving has progressed well. Electric Level 4 connected automated vehicles (CAVs) are now allowed to operate in geofenced areas in several cities around the world as robotaxis and in microtransit demonstration services. Private and public sector interest groups are keen on knowing the timing estimate for CAV use without regulatory constraints. Research reported in this paper contributes information on the likelihood of Level 4 CAV adoption for personal mobility and in shared mobility robotaxi and microtransit services in the 2030–2035 service context.

The methodological framework and its constituent methods are well suited for addressing the uncertainties in predicting the effectiveness of CAV applications and modelling adoption decisions under uncertain states of technology and infrastructure readiness. The Bayesian model also enables the quantification of uncertainty reduction in the CAV adoption decision with demonstration studies.

In the 2030–2035 application context, CAVs are likely to be adopted, provided that the trajectory of progress in technology and infrastructure readiness continues and potential adopters including travelers are offered ample learning opportunities based on high-reliability demonstrations in real life conditions. The threshold level probability of adoption improves significantly with the availability of high-reliability demonstration results to reduce uncertainties in adoption decisions.

In relative terms, higher user and societal gains are obtainable from shared mobility applications than from CAV use for personal mobility. The need for new information to reduce uncertainties in adopting a CAV for personal mobility is higher than CAV applications in shared mobility services. The robotaxi service is likely to require higher trust from potential users than the microtransit service.

These results are consistent with potential adopters’ “need to test it personally”. These highlight the importance of shared mobility service demonstrations and the availability of demonstration vehicles that potential buyers can test for gaining trust in technology.

The products of this research can be used by private and public sector interest groups to enhance technology and infrastructure readiness for Level 4 CAV applications, including the design and implementation of demonstration studies. Another contribution of this research is the methodological framework which offers flexibility to users (i.e., individuals, enterprises) to input their own values of variables sourced from technological and infrastructure readiness forecasts and obtain answers on the likelihood of CAV adoption. Their experience in using the methods illustrated in this paper is likely to contribute knowledge in the technological forecasting field.