• car sharing
• city science
• differential evolution
• genetic algorithm
• geographical
• hotspots
• Istanbul
• machine learning
• optimization
• series
• spatial
• stigmergic
• stigmergy
• temporal
• time

Istanbul is a city in continuous development and this is evidenced by the use of cutting-edge technologies to support various sectors of city management, as logistic.

One of them is car sharing: a company distributes around the city rental cars, that are owned by company, but they can be rent by people for the necessary time. So, if people who live in that area don’t own any car, they can use the nearest shared cars upon reservation for the time they need, then they live the cars to other people that in turn can rent them when they are available.

Logistic managers are interested in providing an efficient car sharing service, but to do that it is important to find the most suitable areas of the cities and, once they have been found, to understand how to distribute efficiently resources to provide a good service.
The suitability of an area can depend on the quantity of people occupy it, the type of prevailing activities or the dynamics in the area.

In this case, suitability is given by the dynamics in the area: it is intended as the variation of quantity of people in a given place and in a given time period.
Given that there is no way to know exactly how many people occupy a given area, this quantity is approximated by the number of calls started in that area and during a certain time period.
This information is retrieved by dataset containing the number of telephone calls started from users through the TURK TELECOM antennas in Istanbul in 2017.

So, the objective is to find hotspots, that are areas of the city characterized by a given dynamic, that is a particular variation in the presence of people over time

To automatically find requested hotspots from this dataset, we will use machine learning techniques to teach a program which are the main characteristics of a hotspot with respect to their presence dynamics, so that it can find unseen ones.

Given that these dynamics have both spatial and temporal components, stigmergic techniques will be used to represent them in a suitable way such that it is easier for the system to learn requested information.
Stigmergy has been used widely in these scenarios: they allow to cluster information considering also how events take place in time, differently from normal clustering algorithms that know how data are spatially allocated.

Until now, the programs that process data through stigmergic techniques have been realized in MATLAB: it is a powerful tool and programming language that can process information efficiently.
The problem is that the actual version of this software processes information sequentially: it could take much time in case of a huge datasets. Moreover, nowadays computers are provided with multicore processors and neural networks are typically trained in super servers with a lot of cores: it would be great to use many of them and not just one.

So, in this thesis I will develop through Python language a software that takes advantage of parallel processing, it uses stigmergy and it integrates genetic algorithms to automatically recognize hotspots and micro-behaviors in them.