key: cord-0765391-m1atpejp authors: Polanco, Liliana Duran; Siller, Mario title: Crowd management COVID-19 date: 2021-04-12 journal: Annu Rev Control DOI: 10.1016/j.arcontrol.2021.04.006 sha: 32108b2acad165270ec0d06918ef24a6516608d0 doc_id: 765391 cord_uid: m1atpejp Crowds are a source of transmission in the COVID-19 spread. Contention and mitigation measures have focused on reducing people’s mass gathering. Such efforts have led to a drop in the economy. The application of a vaccine at a world level represents a grand challenge for humanity, and it is not likely to accomplish even within months. In the meantime, we still need tools to allow the people integration into their regular routines reducing the risk of infection. In this context, this paper presents a solution for crowd management. The aim is to monitor and manage crowd levels in interior places or point-of-interests (POI), particularly shopping centers or stores. The solution is based on a POI recommendation system that suggests the nearest safe options upon request of a particular POI to visit by the user. In this sense, it recommends places near the user location with the least estimated crowd. The recommendation algorithm uses a top-K approach and behavioral game theory to predict the user’s choice and estimate the crowd level for the requested POI. To evaluate the efficiency of this technological intervention in terms of the potential number of contacts of possible COVID-19 infections and the recommendation quality, we have developed an agent-based model (ABM). The adoption level of new technologies can be related to the end-user experience and trust in such technologies. As the end-user follows a recommendation that leads to uncrowded places, both the end-user experience and trust increased. We study and model this process using the OCEAN model of personality. The results from the studied scenarios showed that the proposed solution is widely adopted by the agents, as the trust factor increased from 0.5 (initial set value) to 0.76. In terms of crowd level, these are effectively managed and reduced on average by 40%. The mobility contacts were reduced by 40%, decreasing the risk of COVID-19 infection. An APP has been designed to support the described crowd management and contact tracing functionality. This APP is available on GitHub. In December 2019, in the city of Wuhan, was identified the SARS-CoV-2 virus for the first time. This virus produces COVID-19 disease. The virus spreads from person to person through droplets flying off from sick individuals [1] . To avoid contagion and further consequences, the World Health Organiza-5 tion (WHO) took actions and published recommendations within a short time [2] , but despite those and the global state of alarm, on March 11, 2020, was declared a pandemic. As a prevention and containment measure, most countries have imposed a quarantine on the population. However, so far, the numbers from the WHO show a total of 216 countries, areas, or territories with active 10 cases and more than 1 million deaths [3] . The massive outbreak of the disease has prompted the work of researchers and organizations to create vaccines, detection mechanisms, treatment strategies, containment, and mitigation measures, and decision-making tools [4] . The spread of COVID-19 has caused a sharp drop in economic activity 15 into the control of the pandemic, the containment measures have been relaxed 20 in different ways by the health authorities. Even as restrictions on access to the previous sectors have eased, there will be those who may still avoid visiting these organizations. Currently, masks, antibacterial gel dispensers, disinfectant mats, and temperature controls, among others, have become part of our daily life; nevertheless, these measures do not exempt the danger of being in a crowd. 25 Control strategies for the evolution of the COVID-19 pandemic have been studied by [6, 7] , under the framework of Model Predictive Control, both studies investigated an optimal solution to control the pandemic using dynamic models. The first research work focuses on interventions in a multi-region scenario while the latter on social isolation measures. Both approaches emphasize regional 30 control of the pandemic utilizing isolation measures such as telework, lockdown, or social distance. However, these models are not suitable for individual-level decision-making, which can help to study and prevent dangerous behaviors in society. The early resilience observed in different places to some effects of the pan-35 demic can be explained by the widespread access to digital technologies and the level of digital inclusion. Apps, for example, can be considered as an accessible technology and have been very successful and even institutionalized by governments. A large number of Apps related to the pandemic have been deployed, nevertheless, there are associated with ethical, social, and legal issues that still 40 need to be addressed [8] . Some of the COVID Apps include: Chatbots (Hispabot COVID-19, Victoria, etc.) [9, 10] , Self-assessment and remote diagnosis [11, 12, 13, 14] , Contact tracing [15, 16, 17, 18] , Geolocation [11, 19] , etc. Recommendation systems provide suggestions that allow users with better decision-making. These systems are developed with different approaches and laborative systems, those based on knowledge do not require massive or inclusive private information from users. Consequently, they are considered non-invasive tools. The WHO guidelines "Coronavirus disease advice for the public" [2] , state that to reduce COVID-19 infection rates, the number of people 55 per m 3 (cubic meter) must not exceed 1. The previous restriction directly affects the retail sector in terms of the number of customers that can be concurrently present in a given interior space. It also impacts the consumers as they need to plan and decide which place to visit for shopping or other activities. This consideration applies to a greater or lesser extent when visiting any interior space 60 during the COVID-19 pandemic. In most cases, authorities rely both on the retailers to regulate the number of people (crowd management) in their places and the general public in deciding to visit only uncrowded or low-dense places. Under the current economic crisis, better crowd management regulations and implementations will allow a faster recovery of heavily affected face-to-face de-65 pendent industries and better management of the epidemic (lower infection rate, etc.). To address this hypothesis, we propose a knowledge-based recommendation system and an agent-based model. The recommendation system considers points-of-interest (POI) as recommendation items and suggests uncrowded safe places to visit. It is accessed through an App and intends to provide a non-70 invasive tool. Using a behavioral game theory approach allows to predict the user choice and estimate the crowd level at every POI. We study user behavior through an agent-based model considering individual aspects such as consciousness, agreeableness, openness, and trust. The model also accounts for retail attributes such as capacity and crowd density. The agent-based simulation has 75 two objectives: be an evaluation tool for the recommender system and a means to measure the effectiveness of crowd management interventions at the community level during the COVID-19 pandemic. The work divides into five sections. Section II reviews the state of the art and the development of the agent-based model. Section IV presents the results obtained from the simulation and analysis of the results. Section V contains a 24] , and heterogeneous information [25] . Others focus on the context integration to improve the quality of recommendations as in [26, 27] . While others address specific problems, for instance, in [28] neural networks are applied to generate a sequence of recommendations, or in [29] the authors seek to diversify the POI recommendations to include new places and those that are not often visited. The retrieved data requires a sentiment analysis by sentence to extract char-110 acteristics and ratings of every analyzed POI. Then a clustering algorithm is executed. At the end of this stage, similar POIs are grouped. A vector with K-characteristics is extracted from each cluster. Every user also has a characteristics vector. A similitude metric is then calculated using both vectors. The user is then associated with a cluster and the nearest POI within that cluster 115 to the user location is obtained. For our work, the relevant aspect is the use of a characteristics vector to represent every POI and the clustering approach. The crowdsourcing process can be seen as a way to obtain knowledge, in this sense, we do not have a collaborative recommender instead a knowledge-based recommender updated regularly to keep it up with the new user opinions. Our 120 approach considers a knowledge-based recommendation system with a knowledge base updated through behavioral game theory predicting the user choice. A similarity metric for POI is not needed. In our work, we cluster the POIs according to their geographic distance, in this way, the clustering algorithm does not need to be executed regularly since the location is not likely to change. The model in [26] associates the user interests (obtained from their interactions on the web) with information about places obtained through OSM (Open Street Map) and presents the user with recommendations based on his current 130 location. The system has offline and online modules. The offline module is in charge of computationally demanding tasks such as obtaining POIs from OSM and classify them, as well as calculating the similarity between users and interests. The online module generates recommendations for places according to the user distance and the similarity score between users with similar interests. The relevant aspects for our proposal are the assumed relationship between web interactions and interests, the offline calculation, and that it receives as input the current location of a user. In our work, we assume that the user interests are uncrowded and near places, for this reason, a similarity metric is not necessary. One of our objectives is to predict the user choice, which does not tackle this 140 work. User privacy is an utmost aspect in the design of systems. In [32] we have a model that prioritizes it. In this model, the user and the central server do not and ranks. An algorithm inside the device identifies the group to which the user belongs. From the server is retrieved a list with the best-scored POI within that group. The design based on maintaining the privacy of the users, the clustering technique, and the top-k approach represent the relevant aspects to our work. Even though this recommender system is not collaborative, it might suffer the 155 cold-start issue because it needs existing data to characterize user groups. In our work, we have chosen to cluster POI according to a distance metric, and have assumed the user interest avoiding the cold-start issue. POI recommendations in tourism have been improved by the use of Contextual information. In [33] we have a Context-Aware Recommender System (CARS). This kind of system provides adapted recommendations according to a contextual situation such as weather, season, or schedule. The design is addressed by a context-aware matrix factorization that pairs POI with contextual are predicted according to their personality. Using the Five-Item Personality Inventory in the registration process the system evaluates the personality and 170 can predict user ratings to overcome the cold-start. In this system, the weather provides the context. During a pandemic, different factors can determine when it is appropriate to visit a POI. In the COVID-19 pandemic, we avoid crowded places. Our recommender system assumes the COVID-19 pandemic context and suggests uncrowded places. The prediction of user choice, in our work, is made 175 through logit level-k avoiding the task, sometimes skipped, to answer a test for a user. Following with CARS, the work in [27] presents a recommender system that 180 utilizes a Markov chain to predict contextual information and recommend places that can be visited at the next interval of time. In this case, contextual information refers to crowdedness. The Markov chain has three states: Not Crowded, Moderate Crowdedness, and Crowded. The recommendation process first infers user interest and estimates the crowdedness level from user check-in data. Then 185 an algorithm called Learning-Based Random is applied to score places, the topk best-scored places are given as recommendations. The crowd level estimation represents a relevant aspect for our work, but in the current pandemic context, the three states are not as representative as we need. The data is obtained from user check-in records which are not coherent with our privacy-preserving goal. Utility theory can be useful to recommend POI due to the multivariate attributes. The work presented by [34] is an example of this approach. The utility-based model proposed can learn aspects of the user's preference to pro-from reviews and ratings. With the learned aspects a utility function is calculated. The online module allows users to interact with the system providing the recommendation through a request. The retrieved POIs are the top-k with the 200 highest utility. The applicable aspect in our work is the use of a utility function. Learning the relevant attributes for a POI is a demanding task. We do not need to learn these aspects to evaluate a POI because these are assumed due to the context of the COVID-19 pandemic. We as well use the request model to provide the recommendations. Finally, in our proposal, the execution of the logit 205 level-k model helps to feed the attributes in our utility function, allowing us to supply better recommendations. • Provide reliable information: chatbots implemented by the WHO [36, 37, 38] and by different countries such as Spain [10] and Mexico [9] . • Self-assesment and remote diagnosis: As a remote diagnosis example, "Self-quarantine & safety protection" [11] from South Korea, allows 215 practitioners to determine when to test and monitoring a quarantined patient. Self-assessment apps usually consist of a questionnaire that users carry out to determine if they present symptoms and provide recommendations. This type of application has been implemented in many countries [12, 13, 14] . • Geolocation: Apps made with this purpose handle sensitive data, for that reason, just a few examples can be found. Once that a patient has tested positive "Self-quarantine & safety protection" from South Korea [11] allows to geolocate and monitor; Hong Kong's "StayHomeSafe" [19] works in conjunction with a bracelet, through geofencing it monitors those 225 who have been quarantined. 9 J o u r n a l P r e -p r o o f • Contact-tracing: these type of applications have become popular despite privacy concerns. The first to emerge was "TraceTogether" in Singapore [16] , through Bluetooth it collects information from nearby devices. Following the same approach, other apps have emerged in different coun-230 tries such as Australia [16] , Canada [15] , and Spain [17] . A different approach was taken by Israel with "HaMagen" [18] , which uses geolocation. Check-in applications have emerged as well with this purpose as examples we can find "Territory Check-In" from Australia [39] , "NHS COVID-19" from UK [40], and "NZ COVID Tracer" from New Zeland 235 [41] . Still unfinished but with an ambitious goal is the PEPP-PT (Pan-European Privacy-Preserving Proximity Tracing) [42] project that seeks to track throughout the European Union while maintaining privacy. The agents in the simulation model move around the city to make their purchases. Agents have an app that provides them with the least crowded places 250 near them. The app is responsible for making requests to a server. The server was developed in Python and is responsible for calling the recommendation algorithm. The recommendation algorithm is knowledge-based since the product is known. In other words, grocery store locations are known and the user desired characteristics for such POIs: 1) closeness; and 2) no crowds. Privacy 255 is preserved given that the system does not monitor users to estimate crowds. Rather, it uses behavioral game theory to assess the level of the crowd at each POI. The proposed ABM aims to evaluate the quality of recommendations and 260 the intervention of the recommendation system as a decision-making tool under the COVID-19 pandemic. The model depicts the adoption process of the recommendation system in the form of an App. ABM has been used to study the adoption process of new technology in fields such as smart energy [44] , health systems [45] , business model [46] , and telecommunications [47] . The adoption 265 process in this work is based on the dynamic trust model of [48] . Users develop trust in the App through experience. Agents gain experience as they follow the App recommendation. Trust is calculated following the indirect experience model from [49] . The proposal is based on the following hypothesis: major trust implies that users will make informed decisions, which results in better crowd Each factor is bipolar [52] and address the agent personality in Agent-320 Based simulations. Beliefs: uncongestioned (represents the belief that a place is not crowded), congestioned (represents the belief that a place is crowded), end shopping (represents that an agent has found everything he needed). Desires: need supplies (represent the need to do the shopping). Emotion: f ear (Emotion fired when an agent perceives a multitude in the store he is heading to, an agent can perceive a multitude in a 50m radius). • Person: A species that inherits from People and represents an agent that has no access to the recommendation system. • User: A species that inherits from People and represents an agent that has access to the recommendation system through the App. This species contains particular attributes. The Initialization of the model takes into account the following considerations: (i) the infrastructure entities such as the streets, residential and commercial blocks are created from GIS files (shapefiles); (ii) the store entities are loaded from a CSV file that provides location coordinates, store name, and ca-395 pacity; (iii) a simulation parameter sets the populations of human agents; and (iv) the allowed crowd percentage is specified using a simulation parameter that can be changed at execution time. if !has emotion(f ear) then 5: selected ← one of (knowledge base)) 6: The influence diagram " Fig. 4 " describes the agent s decision-making pro-400 cess. The Choose Store box represents the decision that agents must be make. Agents have two options that will lead them to a store: the App or their an influencing factor. The influencing factor represents how much the experiences of others affect our trust. According to [55] obedience is a way of social influence, we consider obedience as the influencing factor that represents how much we consider the evaluations of others. Obedience according to [55] can be predicted by personality. The agreeableness and conscientiousness personality 415 traits predict our level of obedience. T rust is developed through experience but the willingness to experiment is determined by the personality of each individual, in the dynamic trust model it is called the flexibility factor. The openness personality trait is related to our willingness to experiment [51] , a value close to one is interpreted as a greater inclination to have new experiences. In our 420 model, openness represents the flexibility factor or the willingness of an individual to experiment with the App. When trust has been computed, the agent randomly decides whether to use the App, this random choice depends on the trust level, with a higher trust level it will be more likely that the agent will choose to follow the App. App suggestions are computed using the distance 425 and the estimated crowd at every POI, this process is described in section 3.2. When the agent chooses to follow the App recommendations, they will select the one with the best score. When the agent decides to use his knowledge (de- The indirect experience E i , in this work, refers to an average of the current trust of all agents (except the one that is currently computing) and is calculated by: where N represents the total number of agents, in this case of the User type, T B the current trust of each User agent. In our model, the direct experience E d for each agent is calculated by the number of good recommendation with respect to the total amount of followed recommendations: A recommendation is considered a GoodRecommendation when it leads the agent to a location with a crowd percentage lower than the belief congestion J o u r n a l P r e -p r o o f attribute. The overall experience E to calculate the trust is computed by: where α A represents the willingness of the agent to be influenced by the experience of others. [56] suggests that behavior is influenced by what we think others approve. Our model considers indirect experience E i A as a measure of approval. According to [55] obedience is a way of social influence. We consider obedience as α A in the trust model. E d A states the direct experience of the agent. Finally, the trust is computed by: where T A represents the current trust of the agent, E A the overall experience, 435 and γ a personal characteristic of flexibility. The factor γ in this model is associated with the openness aspect of personality since it is linked to our proneness to accept new experiences [51] . The App gets recommendations based on the distance, and the estimated crowd level, a detailed explanation of this process is provided in the next section. The authors in [29] state that a POI recommendation problem is equivalent to a k-POI selection problem. In this sense, a recommender system selects the first k POIs that match user preferences from a set of candidate POIs. Using this approach, we have developed a knowledge-based recommendation system. 445 We define the general characteristics based on the classification framework proposed by [57] for knowledge-based recommendation systems, see "Tab. 1". The system takes the POI database and executes k-means as a clustering algorithm. POI clusters and mean points are saved in the database. The recommendation algorithm takes the mean points to classify the location of a user to a cluster. For every POI within a group, it obtains the coordinates, capacity, allowed percentage, and currently estimated multitude. With the retrieved data, the system calculates a payoff function S based on the distance to the user Degree of automation to generate user profile High, users do not define their profile with actions. The system assumes that they want to find the nearest and least crowded places to their location. Provide the K least crowded places and nearest to a given location. The penalization works as a bonus b given to those POI with a minor estimated multitude than the multitude threshold. The payoff S is computed using three weights: distance weighing W d , multitude weighing W m , and bonus weighing 21 J o u r n a l P r e -p r o o f W b . The best POI to recommend is the one with the biggest payoff score. Once the payoff is calculated, and the k-best POI selected, logit level-k calculates the probability of a user to choose each place, updating the estimated multitude at the POI with the higher probability. The logit level-k model is 450 computed using the following equation [58] : where s −i denotes the assumed strategy profile of other agents, k represents the rationality depth parameter, λ the precision parameter that controls the liability to utility differences, a i represents the selected recommendation and U i references to a utility function. We consider a depth k = 1 because it has 455 proven to be a representative level for human agents [59] and has been applied in similar situations [60] . a i represents a utility value given when selecting a recommendation and it is defined as follows: where S i represents the payoff associated to the selected recommendation. which represents the choice of the opponent's assuming that his level is 0, which means that it is chosen randomly. Finally the utility is computed by: The recommendation algorithm is shown in " Fig. 5 ". • The simulation runs in one day, for example, Saturday, which is the preferred day for shopping. • The population of agents represents the people who make purchases during a day. • The stores do not have a monitoring mechanism at the entrance because we are exploring how the tool by itself can prevent multitudes. • Trust evolves following the indirect experience model from [49] . Agents 470 gain experience by using app recommendations. • The quality of recommendations is the factor by which the trust is increased. A good recommendation is one that leads to a place perceived as uncrowded, otherwise it is considered a bad recommendation (place perceived as crowded). • Personality plays a role in the decision-making process and trust evolution. Openness is linked to our readiness to have new experiences (equivalent to γ in the trust model). Agreeableness and Conscientiousness are linked to obedience [55] , we considered obedience as how others influence us (equivalent to α in the trust model). • Trust is a measure within the interval [0,1]. Initial trust is 0.5, this comes from the fact that we assumed that our App is institutionalized, which gives more credibility to the tool. • A track of mobility contacts is made as agents move through the map. A mobility contact is made when two agents violate social distance guideline • Our experiments were executed using a zone from Guadalajara city in Mexico, the area around the Galerias Mall. The selected zone is highly commercial and surrounded by different residential areas (see " Fig. 14" ). • The selected stores are the biggest in the zone (Galerias, Walmart, SAMS, Costco, Comercial, and Chedraui) • A population of 1000 agents was set for exploration purposes. • A 25% of the total capacity is allowed at each store. However, the stores do not have a monitoring mechanism at the entrance because we are exploring were even but as the simulation unfolds, the App recommendations were preferred (see " Fig. 9") . The average trust in the application increased from 0.5 to 0.77. The observed track of mobility contacts changed in every scenario. We can appreciate that scenario 1 reached 42322 contacts, while the others presented an improvement reducing the contact to 34484 and 25124, re-525 spectively (see " Fig. 13") . For a summary of the presented peak values and the number of times that the allowed capacity was passed, see "Tab. 2". It is possible to appreciate how in every scenario the peak distribution is more balanced allowing lower peak values. The balanced distribution affects the number of times that the allowed capacity was surpassed, this To address this, we started from the fact that, currently, crowd management measures commonly include control strategies such a temperature measurement, the use of disinfectant mats and hands sanitizer. Every public or private space that allows the gathering of people must carry out these preventive measures, • Anonymous registration. To use the App, the user register anonymously. A hash code is generated at the beginning and no more personal data is required. • Record of temperature measures. Temperature measurement is a control strategy in the current pandemic. The App can keep a track of all the temperature measures made at every visited place. • Notify infection. When a user is infected he chooses to share the history of all visited places, then the App sends an alarm to all the users that could had been in contact with the infected user. • Heatmap to show crowded places. Through the heatmap, the user can visualize which places are more convenient to visit. • Recommend convenient places. Besides the heatmap, the App can suggest near and uncrowded places. Key features for establishment: • Crowd level monitoring. Keep track of all the check-in/out to the establishment, allowing continuous monitoring to not incur in a fault to 575 mitigation policies imposed by health authorities. • Check-in. The user registration process returns a QR code. The code is unique and should be scanned at every establishment to register the check-in and maintain the temperature records. • Notification when allowed capacity is reached. The App code is available on GitHub. New functionality is being currently added and access can be granted upon request to the authors. Some screenshots are presented in figures: " Fig. 15", "Fig. 16" and "Fig. 17 ". On the right, check-in screen. In this work, we have presented a technological intervention in the context of as successful as we thought, however, we can observe improvements in the peak 590 levels reached in every scenario. For instance, scenarios with the intervention never reach 60% of capacity. In every simulation we can observe that some stores remained more crowded than others, this can be due to the geographic location. In terms of mobility contacts, it varies in every scenario. It can be seen that the higher the access to the recommendation system, the lower the 595 number of contacts produced. The proposed model can be complemented with SIR, SIS, or SEIR models to further study the spread of the coronavirus during a contact situation. Finally, it is possible to estimate and manage the crowd level at POIs using little information and maintaining user privacy. A familial cluster of pneumonia associated with the 2019 novel coronavirus indicating person-to-person 615 transmission: a study of a family cluster Who coronavirus disease (covid-19) dashboard -who coronavirus disease (covid-19) dashboard A comprehensive review 625 of the COVID-19 pandemic and the role of IoT, drones, AI, blockchain, and 5g in managing its impact Estimating the economic impact of the covid-19 shock Model predictive control to mitigate the COVID-19 outbreak in a multi-region scenario An optimal predictive control strategy for COVID-19 (SARS-CoV-2) social distancing policies in brazil The institutionalisation of digital public health: Lessons learned from the COVID-19 app La adip y el gobierno de la ciudad presentan victoria, el robot que colaborará en la atención de la ciudadanía Hispabot covid-19: resuelve tus dudas sobre el coronavirus Quarantine-report • self-check -apps on google play Asistencia covid-19 -apps en google play Covid-19mx -apps en google play Covid-19 situation update Download the covid alert mobile app to protect yourself and your community -covid-19 (coronavirus) in ontario COVID-19 apps in singapore and australia: reimagining healthy nations with digital technology Hamagen -the ministry of health app for fighting the spread of coronavirus Covid-19 thematic website -together, we fight the virus -"stayhomesafe" mobile app user guide Tourism knowledge discovery in social multimedia A survey of point-of-interest recommendation in location-based social networks Point-of-interest recommendation in location based social networks with topic and location awareness SIAM International Conference on Data Mining Exploiting the roles of aspects in personalized POI recommender systems Point-of-interest recommendation using temporal 700 orientations of users and locations, in: Database Systems for Advanced Applications Towards real-time demand-aware se-705 quential POI recommendation A novel point of interest (POI) location based recommender system utilizing user location and web 710 interactions A markov 715 chain model for integrating context in recommender systems Attention-based recurrent neural network for 720 location recommendation Geographic diversification of recommended POIs in 725 frequently visited areas Using function approximation for personalized point-of-interest recommendation Cold-start point-of-interest recommendation through crowdsourcing Privacy-preserving POI recommendation using nonnegative matrix factorization User personality and the new user problem in a context-aware point of interest recommender system MARS: A multi-aspect recommender system for point-of-interest Tecnologías para luchar contra la pandemia covid-19: geolocalización, rastreo, big data, SIG, inteligencia artificial y privacidad, El profesional de la informacióndoi Who launches a chatbot on facebook messenger to combat covid-19 misinformation Who health alert brings covid-19 facts to billions via whatsapp The territory check in -coronavirus (covid-19) Nz covid tracer -apps en google play Github -pepp-pt/pepp-pt-documentation: Documentation for pan-775 european privacy-preserving proximity tracing Building, composing and experimenting complex spatial models with the GAMA platform Agentbased simulation design for technology adoption Agent-based modeling and simulation to adoption process of information technologies in health systems Proceedings of the Second International Conference on Internet of things, Data and Cloud Com-795 puting Simulation-based prediction for 5g mobile adoption Repairing trust in an e-commerce and security context: an agent-based modeling approach Agent-based and population-based modeling of trust dynamics Railsback, The ODD protocol: A review and first update An alternative "description of personality": The big-five factor structure Modeling the robo-advisor ecosystem: 820 Insights from a simulation study Grocery shopping behavior and the pandemic -deloitte insights Personality predicts obedience in a milgram paradigm Using social and behavioural science to support COVID-19 pandemic Knowledge-based recom-840 mendation systems Decision making under uncertainty : theory and application On the beliefs off the path: Equilibrium refinement due to quantal response and level-k nobody goes there anymore -it's too crowded': Level-k think-850 ing in the restaurant game