In this work, we explore the ride-hailing assignment problem for the city of San Francisco, USA. We aimed to optimize drivers' behaviors for simultaneous objectives such as maximizing service level, minimizing CO2 emissions, and minimizing riders' waiting times. We tested four different policies with a real-world case study that involved GIS and GPS data. We developed, in Python, a realistic simulation model that captured the most relevant characteristics of a ride-hailing system; moreover, we used the current road networks to approximate real-time road conditions. Furthermore, it is the first work that tests the effects of different passengers' arrival conditions and analyzes the multiple objectives for different zones of a large city. Results suggest that passengers' arrival conditions affect the four proposed policies nearly identically. Besides, the policy of drivers remaining static (in place) instead of driving while searching for passengers had the highest service level and lowest average distance per ride. Even though it could sound unrealistic under current conditions, this behavior could be achieved soon when all ride-hailing services are serviced by self-driving cars. This work will be presented at the 2021 Winter Simulation Conference.

During the first days of COVID-19, every media outlet used the expression 'flattening the curve'. Since coronavirus is a socially transmitted disease, the urgency to slow its spread meant that local government had to impose restrictive measurements; however, the layman did not necessarily understand why these measures were in place. The short answer was that slowing the spread had a direct impact on lowering the number of cases requiring ICU beds at once, which meant a reduction in the number of deaths from the virus or other causes. We did this educational project to explain two important points: i) mobility patterns have a significant impact on how the disease is spread geographically, and ii) explain that in a Susceptible-Infected-Recovered (SIR) model a tall and skinny curve is bad, and must be flattened if the goal is to save lives.

For the first goal, the video below shows that for the Province of Concepción, Chile, mobility patterns (from a pre-COVID origin-destination matrix) have an important effect on the Disease spread, susceptibility, and infection intensity. It implies that contiguity, in this case, is not only governed by physical distances but also by travel patterns. Additionally, considering the socio-economic dimension of the problem (the figure above), efficient handling of the pandemic becomes difficult and less intuitive. To model this regional contagion dynamic we implemented a straightforward procedure presented in this Nature article.

For the second objective, we tested three different scenarios: 1) pre-pandemic mobility patterns (the same as in objective i), 2) Reduced mobility patterns, computed by experts after a curfew and mobility restriction were imposed in the region, and 3) the same mobility patterns as in 2, but with the early reopening for all areas when the number of contagions was below a certain threshold. The video below shows graphically the effects of these policies on "flattening the curve", and the importance of social collaboration to achieve the goal of reducing the number of deaths due to the virus.

We propose a new survival-based real-time optimization framework that maximizes the survival probability of patients suffering from life-threatening emergencies while maintaining a balanced workload among the crew. Our approach embeds a mixed-integer linear programming (MILP) optimization model inside of a discrete event simulation (DES) model (programmed in Python). The DES model, after a predefined number of events, feeds the MILP model with real-time vehicle utilization metrics to maintain an optimal overall relocation policy. Moreover, we evaluate the impact on the system’s congestion of a mixed dispatched policy that includes the use of external ride-hailing services for minor emergencies. Our case study in New York City indicates that the proposed model increases the average percentage of life-threatening emergencies responded in less than four minutes by more than 10% when compared to a coverage model from the literature. Besides, when compared to a static version of the proposed model, the online approach improved the number of life-threatening emergencies responded in less than six minutes in 14% overall and more than 17% for the four minutes threshold. Finally, the introduction of ride-hailing services presents an important opportunity for reducing personnel workload by 13.5%, on average, and operational costs in $141,220 per day of operation for the entire system. This work is currently under review at the OMEGA journal.

When COVID-19 hit the Bio-Bio region, Chile, it affected all dimensions of people's daily life. One challenge that arose when schools switched to online learning was the handling and food delivery to students (now at home) with access to meals through federal school meal programs. At the time, the local government had different ideas about how the delivery process should be, and they wanted to test them to select the most promising approach. We implemented a simulation model that, based on GIS data, evaluated the food delivery process following their proposed strategies and some others we suggested.

Successful tsunami evacuation drills are critical to saving lives. However, city-wide site emergency simulations are not only difficult to carry out but also expensive. In this work, we implemented a realistic city-wide agent-based simulation that uses a synthetic population, geographic information systems (GIS), and origin-destination data to study evacuation routes that minimize the overall evacuation time. This topic has a great impact on tsunamis-prone countries, such as Chile, that suffer periodic natural disasters. Our "in-silico" representation of the city of Antofagasta, Chile, also includes street elevations to penalize evacuees' escape velocity. We evaluate four scenarios: i) every family evacuates to a meeting point; ii) a stochastic escape scenario, in which, each family randomly decides either to escape to a building (vertical evacuation) or a meeting point; iii) assign each family to either a meeting point or a building, using an assignment optimization model, and iv) a scenario that studies the effect of city-wide mobility (using origin-destination matrices) on the evacuation times.

Vehicle sharing systems are a new mode of transportation that has spread worldwide widely. Bike-sharing and scooter-sharing appear as solutions to the last-mile problem and environmentally friendly alternatives to private cars. However, due to asymmetry of demand, these systems tend to suffer severe unbalance. In this work, we develop a realistic discrete-event simulator designed to test a user based rebalancing strategy using a reinforcement learning approach. We modeled the Spatio-temporal behaviour of a vehicle-sharing system by combining an inhomogeneous Poisson process, a Spatio-temporal Kernel Density Estimator and a conditional Kernel Density Estimator. Next, we use our simulator to train a reinforcement-learning algorithm called Hierarchical Reinforcement Pricing (HRP), which aims to rebalance the system by giving price incentives to users to influence users’ route trips. It is worth mentioning that HRP was initially proposed for bike-sharing systems. Our results suggest that the HRP algorithm does not perform well in such a stochastic environment, reaching a Decrease Service Ratio (DUR) of 12% on average under a maximum incentive of 3$. Although part of the problem is due to the poor performance of reinforcement learning algorithms in stochastic environments, the differences in users' behaviour compared to bike-sharing systems are also critical. Several improvements could be made to enhance the performance of HRP, using the advances in continuous control tasks. Finally, our simulator could be used as a basis to test reinforcement-learning user-based strategies under realistic scenarios. A visualization of our model can be found at this link

Here we tackled a real-life problem faced by Chilean National Firefighters' Agency. We introduce the Facility Location and Equipment Emplacement Technique with Expected Covering (FLEET-EXC) model, an emergency facility location model that maximizes the coverage of expected demand. We consider multiple regions, demand types, vehicle types, and region-dependent dispatching rules. Our solution approach is mixed-integer linear programming (MILP) model that considers vehicles' average utilization to compute expected demand coverage, coupled with a Hypercube Queueing Model that provides the aforementioned utilization. The goal of this procedure is to update parameters until the resulting vehicle utilization from the optimal solution is the same as the ones used to compute the expected demand of the MILP model. The synergy produced by relocation and location of new facilities notoriously improves the emergency coverage, providing insights for strategic decision making. This work is published in the Computers & Industrial Engineering journal.

In this project, we continued our work with the Chilean National Firefighters' Agency regarding the study of the strategic decision-making for proper location of fire stations as well as assignment of vehicles to each station with the goal of improving emergency response. We implemented an iterative simulation-optimization approach that based on some precomputed utilization parameters updates the optimal location of vehicles and fire stations. For the optimization part, we introduce a robust formulation of the Facility Location and Equipment Emplacement Technique with Expected Coverage (Robust FLEET-EXC) model, which maximizes demand considering vehicles’ utilization. Additionally, the emergencies arrival process is modeled by a spatio-temporal sampling method that loosely couples a Kernel Density Estimator and a non-homogeneous non-renewal arrival process with a Markov-Mixture of Erlangs of Common Order model as a base process. After experimenting with several configurations we show that the model is statistically better than the deterministic FLEET in all but one of the experiments, resulting in up to 6.42% more coverage. This work is published in the Computers & Industrial Engineering journal.

In a tidal power plant, the most important devices are the tidal current turbines that harvest the kinetic energy contained in the tide and convert it into electricity. The number of turbines to install, their spatial configuration, and the connection of submarine cables are critical aspects in the development of marine energy projects. In this project, we proposed a mixed-integer programming methodology to design tidal current farms considering both cost and benefits. Our model solves, concurrently, the turbine’s locations and cable infrastructure by computing the number, locations, submarine connections of tidal turbines, and the tidal current farm’s overall capacity. We consider location-based costs, including cable routing and loss from wake effect and electrical transmission, to maximize the project’s economic viability. We demonstrate the methodology’s utility by determining the optimal turbines’ configuration in a real-life case study in the Chacao Chanel, Chile. Finally, we published this work in the Applied Energy journal.

Over the last few years, wildfires have increased in frequency and intensity all around the world. Wildfire susceptibility mapping with machine learning has been proven useful for fire-prevention plans, turning into an indispensable tool in wildfire prevention. In this project, we demonstrate that an Ensemble model, based on two deep learning networks previously presented in the literature, can achieve remarkable results for forest fire susceptibility. Our model was trained on fire data occurrences in two regions in Chile, from 2013 to 2019. Satellite imagery data for fifteen fire-influencing factors in the study area were retrieved to build a dataset to extract the samples to train the models. These factors include elevation, aspect, surface roughness, slope, minimum and maximum temperature, wind speed, precipitation, actual evapotranspiration, climatic water deficit, NDVI, land cover type, distance to rivers, distance to roads, and distance to urban areas. During training, the best sample patch size was 25 × 25 pixels, and the highest area under the curve (AUC) for the Ensemble model was 0.953. Besides, the Ensemble model achieved the best accuracy, sensitivity, specificity, negative predictive value, and F1 score. Finally, the predicted susceptibility maps suggest that static variables can be considered as predisposing factors, whereas dynamic variables affect the intensity of the predicted probabilities, with an important role of the anthropogenic variables. A complete description of this project can be found in Ecological Informatics.

In this project, we taught the Baxter robot to identify and sort some elements. My student (Felipe Sepulveda) and I experienced firsthand that trivial tasks for humans are still significant challenges for machines. For this reason, we implemented two independent object detectors in the robot. On its left hand, we trained two shape detectors, a Gradient Boosting Classifier and a Multilayer Perceptron Classifier, with data generated using the theory of complex contour analysis. Then, we taught the right hand to use a K-Means algorithm to distinguish colors in a dataset generated using the back-projection theory. In our physical experiments, Baxter observes a collection of objects, randomly positioned within a search space, with the aim of identifying not only the aforementioned space but also recognizing each of the objects to deposit them in their corresponding containers and ignore those that do not possess the properties sought. The Gradient Boosting Classifier gave the best performance on the left hand of the robot, with an f1-score of 94.65 % over a validation set of 925 images. For the color detector (right hand), a silhouette coefficient of 91.68 % shows that K-Means requires a total of 3 clusters over 654 training images. Moreover, because of the theories behind the data generation process, shape features are invariant to rotations, translations, and scaling, while color features are robust to illumination variations. Unknown objects are ignored using properties of the proposed features; thereby, there is no need to add more data to account for unobserved classes during the training process.

In this project, we explore the opportunity of using Machine Learning algorithms to model travel mode choice. Currently, the multinomial logit (MNL) and the mixed multinomial logit (MMNL) model are the most popular algorithms to model the problem. However, machine learning (ML) methods have begun to gain ground in the transportation field, showing a high predictive power that surpasses logit models. In general, studies comparing machine learning methods and logit models mainly focus on predictive accuracy, while others –to a lesser extent– focus on post-hoc explanation analysis. A differencing element of this works was that we shed light on the explanation capability of ML models by computing the effect of different variables not only on the overall prediction but also on the prediction of different choice alternatives using an agnostic-model method. We also present several empirical applications using four different mode choice datasets where our results show that Neural Networks generally perform better than other models in terms of accuracy and interpretation. Moreover, we emphasize the importance of analyzing the equivalence between the models in order to complement the explanations obtained through the two approaches to support management decision-making and to better understand the factors that determine people’s travel behavior.

This work presents a methodology for the segmentation and visualization of academic information of first-year students using machine learning techniques and visualization algorithms. We show that students from high-performance secondary schools have better chances of academic survival during their first year. Besides, they tend to choose either Chemical, Industrial, Civil, or Mechanical Engineering. Conversely, students from the worst-performing groups, comprised mainly of public high schools, enroll in Telecommunications and Materials Science Engineering. We use Self–Organizing Maps to generate a non–linear projection to a two–dimensional space spanned by the students’ academic records. Then, we clustered students applying the k-means algorithm on the resulting projection. Next, we use the Local Interpretable Model–agnostic Explanation (LIME) algorithm to determine the relative contribution of the explanatory variables on each cluster. After that, we geospatially project the clusters to observe patterns among high schools and generate network visualizations to understand the flow dynamics to the programs offered by the College of Engineering at the University of Concepcion. Finally, we use the Squarified Treemap Algorithm to analyze the clusters' composition over time.

This work shows the use of modern machine learning techniques to solve current problems faced by the forestry industry. More specifically, we compare the predictive performance of several algorithms on estimating the dry weight, in tons, of chip residues. The dataset contains samples obtained over 22 months from 220 trucks from 17 farms within the area spanned by the Bio-bío and Maule regions, Chile. Once the trucks arrived at each power plant, samples were collected and dried to compute the dry tons, set as the dependent variable. Using open-source software implementations of state-of-the-art algorithms we determined, for our data, that the non-parametric models Gradient Boosting Machines (GBM) and Neural Networks (NNET) are not statistically superior to the LASSO regression (GLMNET), a penalized version of the Linear Regression model. Additionally, we observed that seasonality affects the ratio of green tons to dry tons a truck can deliver to a power plant during the year. Finally, the continuous variables green tons, elevation, east and north (longitude-latitude) also contribute to explaining the dependent variable. We published this work in the International Journal of Forest Engineering.

Here we had the challenge of finding a relationship among several oceanographic variables and fish catch sizes that could lead a fishing fleet to good fishing zones. We downloaded satellite imagery (from Copernicus) such as seawater velocity and direction, sea surface temperature, plankton concentration, sea height, and salinity. With this information we could visualize ocean currents that fishes use as "highways" during their migration process, being the Humbolt and the Gunther currents the most relevant for our study. With the collected information we also identified places where the sea surface temperature changed abruptly. Finally, using a clustering algorithm we could observe that faster current velocities occur at zones where sea surface temperature changes. Interestingly, when we mapped the movement of fishing boats we observed that Captains tend to sail to exactly the same areas where the aforementioned phenomena occur.

This project is quite interesting because we created a reliable and granular synthetic population to serve as an input for microsimulations of social dynamics. In our case, Chile releases anonymized data, at the 5-zip area code level, of each individual (18 million records approx.), but the country also provides marginalized information at the block level. Then, the problem can be seen as a particular instance of the generalized assignment problem. Based on this, we implemented a mixed-integer programming model that takes advantage of the particularity of the data to generate a synthetic population. The model relies on slack variables to achieve a feasible association between the disaggregated and aggregated data mismatches with the purpose of finding an optimal fit that reproduces the underlying structure of the real population, as close as possible. It does it by assigning households to each of the blocks within a census area (5-zip area code). Likewise, cartographic data was used to generate Lon-Lat coordinates of all households and individuals within their respective blocks. As a case study, the Province of Concepción, Chile, was selected. The results were compared with a heuristic proposed in the literature, showing that our model is superior in fitting the population at the category and area level. Furthermore, the proposed model will find the optimal solution when consistent and adjusted data are available, but when having mismatched data the assignment will not be optimal but it will remain significant from a statistical viewpoint.