Machine learning boosts Shotl users’ experience

The influential professor of Machine Learning, Pedro Domingos, states that “people worry computers will get too smart and take over the world, but the real problem is that they're too stupid and they've already taken over the world”. Intelligent transportation systems are no exception.

Machine Learning (ML) is a field of Artificial Intelligence (AI) which focuses on the development of techniques that allow computers to learn how to perform a specific task. They rely on mathematical models and algorithms that use data samples, known as training data, which in turn provide the information needed to make decisions. 

ML can be divided into two major subfields: supervised and unsupervised learning. Supervised learning is a set of ML models whose aim it is to predict the value of an explicit dependent variable. The training data used in these cases consists of observations of the dependent variable for different values of the other (independent) variables, potentially being a very large data set. 

Supervised learning techniques can be split into regression and classification. In the former, the dependent variable to be predicted is a real number, whereas in the latter it is a class from a finite set. There is no clear border between regression and classification models, as some regression models are often used to predict a class rather than use the specific number predicted by the model. 

On the other hand, unsupervised learning techniques are not intended to predict a specific variable, but rather to give insights regarding a dataset. Clustering algorithms, that create clusters or groups of data without predefined classes, are the most well-known example of these.

How can ML, and more specifically supervised learning and predictive modeling, help improve Shotl users’ experience?

User pickup and dropoff times can be simplistically estimated from vehicle routes by adding up the travel times provided by the routing API, thus based only current services. The main drawback of this approach is that it does not take into account potential services that can enter the system at a future stage.

Shotl’s delay prediction algorithm gives an estimation of both the waiting time at the pickup location and detour/delays due to stops at the pickup/dropoff locations of other ongoing services. The training data here is the historical data of services in a given area, and the algorithm can detect demand patterns that are specific to each day of the week as well as the different time ranges during a single day.

Demand forecasting is interesting from a user planning perspective: if it is likely that there will be a delay in the service, users are informed of the expected pickup and dropoff time and are then able to make informed choices. It is even more advantageous when we look at how demand forecasting can improve the overall quality of service.

When a vehicle finishes its route and all requests have been fulfilled, it is typically sent to the depot to wait for more incoming requests. Idle vehicles can instead be redeployed to areas of high demand to reduce user waiting times. Good prediction of potential short-term demand, as well as the current state of the system, including the positions of vehicles and their pending services if they are not idle, are crucial in Shotl’s vehicle relocation algorithm.

In a nutshell, Shotl uses the data generated by users and drivers to build modern high-performance ML models that enhance the user experience and allow for a more efficient, less costly transportation system.