What do we exactly do in TensorCell?
Research questions and applications
How to optimize road traffic in cities, make it smoother and more environmental friendly?
How to optimize the process of treating cancer?
How to optimize home delivery, make it more efficient and receive your shopping faster?
These (any many other) are some questions we try to answer in our research work. One may think:
"How on Earth you can work on both, optimizing transport and treating cancer? Why do you work on such different topics instead of focusing just on a single topic?!"
Well, the World is a very complex and interesting place and it turns out that such (apparently) different phenomena as controlling road traffic and treating cancer have some similarities and we, scientists, may try to tackle them using the same universal techniques. Moreover, it is possible that some approaches from one area may turn out successful and be an inspiration to solve similar problems from other areas. In our case, it turned out that the traffic signal control approach from traffic management became a successful tool for optimizing the process of radiotherapy and cancer treatment! How did we achieve it? Let's go back to the beginning...
A brief history
A long time ago, in a galaxy far, far away ... a PhD student from the University of Warsaw, Paweł Gora, started working on modelling and optimizing traffic. To do that, he developed Traffic Simulation Framework, a tool for simulating and investigating vehicular traffic in cities. You can see its screenshot below or watch a screencast presenting its functionalities: https://www.youtube.com/watch?v=94RatF5SrLw
Looks cool? This tool found applications in several projects, e.g., it was used to generate data for the traffic prediction contest supported by TomTom: http://tunedit.org/challenge/IEEE-ICDM-2010, but one of its most interesting applications is testing traffic signal control algorithms. Paweł was working on this topic for a couple of years and he achieved good results by using genetic algorithms and high-performance computing platforms, but the biggest challenge was a large time complexity of this approach. Therefore, he started thinking about speeding up those experiments. One of the promising approaches was based on building metamodels (surrogate models) approximating traffic simulation using machine learning. The initial experiments gave great results, were presented at NIPS 2016 conference and were a base for the further research within the TensorCell project.
Later, it turned out that a very similar approach may be (unexpectedly) applied in cancer treatment optimiztion. This research started from the observation that both phenomena (road traffic and cancer growth) can be modelled using (probabilistic) cellular automata, and in both cases the process is governed by some control rules (i.e., traffic signal settings or radiotherapy protocols) which we would like to adjust in order to minimize the outcome (time of waiting on a red signal or the number of cancer cells). Therefore, in both researches (traffic optimization and cancer treatment) we use microscopic simulations based on cellular automata in order to model the complex process (i.e., road traffic or cancer growth), we use genetic algorithms to optimize the given process (i.e., to minimize the total time of waiting on a red signal or to minimize the number of cancer cells), finally, we use metamodels (or surrogate models) approximating outcomes of simulations in order to speed up computations.
Since complex processes are ubiquitous and cellular automata are universal tools for modelling real phenomena, we are also looking for new applications of this approach.
Where we are today?
From 2017, new researchers started joining the TensorCell project. Now, the group is still working on optimizing traffic signal control. We use mostly methods based on machine learning (graph neural networks and LightGBM), metaheuristics, traffic simulations, reinforcement learning. We've published 5 researcher papers on conferences such as NeurIPS, MT-ITS, TFML. Beside traffic signal control, we also work on optimizing the logistics (especially last-mile delivery), more specifically: to solve the so-called Vehicle Routing Problem with Time Windows.
TensorCell is now formed by more than 20 researchers, some of them are professionals, some are scientists, PhD students or students. From the very beginning, we work 100% remotely.
We are now focusing on research but at some point we would like to implement our methods in practice, to solve the problem of inefficient traffic management and to optimize the process of radiotherapy. Therefore, it is possible that at some point the project will evolve into a startup.