My motivation for doing my job, and before that, being a student and aiming at becoming a researcher, is mostly a search for understanding things, and think about things. And sharing this with other people, colleagues, students, whoever might be interested in, together going further towards new unknown territories.
I am not really interested in digging further, rather in finding new areas to dig into. I prefer exploring new areas, rather than exploiting.
This led me to change my field of interest very strongy along the years and visit various remote areas of computer science: symbolic AI, computer architecture, languages and compilers, genetic algorithms, combinatorial optimization, multi-agent systems, machine learning, data mining, numerical/statistical AI.
This also led me to work with researchers of other fields: marine biologists, behaviorist psychologists, agronomists, health researchers.
This also led me to collaborate with companies on various topics to understand companies point of view, needs, etc.
In this section, I describe my current research activities and projects.
Fundamentally, my research domain is machine learning, and particularly sequential decision making under unceertainty, mostly reinforcement learning, a bit of bandits, and a bit of any other relevant topic.
I have decided a couple of years ago to focus on a set of research topics and applications of my research. After 30 years of research activities, I feel like I need to try to make my research useful, at least to some extent, to the well-being of society, humankind, and Life on earth (!). This led me to select the set of research questions I want to investigate in the coming years. Regarding applications, I want to focus on health, and sustainable development.
I think that machine learning has brought new tools, algorithms, methods, ... that have to be used, put to the test, and tailored to important fields of applications. Applications also bring questions and challenges that have to be dealt with by fundamental research, and drawing my attention to issues that are meaningfull.
My main field of research is reinforcement learning and, more generally, sequential decision making under uncertainty. The biggest achievement to date of RL is learning to play go better than any human being: this is indeed an achievement, but this is still very modest when I compare the complexity of go to the complexity of applications I mention above, and applications that people are interested in be it in companies or in other institutions (hospital, ...). To be able to tackle such complex applications, pure RL is far from being enough. We need to combine, or hybrid, RL with all possible techniques and knowledge that can be useful to solve such problems. Hence, today, beyond trying to make RL better, my main research line is how can we do this? Very often, a learning algorithm re-discovers things human beings know perfectly well: obviously, the learning algorithm would benefit from this knowledge in the first place. An other thing is that when one looks at how an RL algorithm is learning to solve a task, one identifies ways in which it could do better if it was able to use even basic rules; based on statistics and correlation, RL algorithm induce regularities but they can not induce logic rules, they can not even use logic rules: they gather thousands, millions, sometimes billions of data but they can not induce even the simplest rule out of them. Once being able to use rules, the RL would be able to reason, anipulate knowledge to deduce new facts: again, such simple things are not done by today RL algorithms.
In this section, I say a few words about the directions in which I work.
To sum-up my thinking about so-called AI, there is absolutely nothing intelligent in it: current AI systems, sometimes achieving remarkable feats, are completely stupid, and rely exclusively on computation power, and the smartness of their designers and implementors. I want to go further than that. There are numerous tasks where current state-of-the-art AI is simply unable to do anyhting or, if it can, does it in a completely stupid way, a completely different way from the one we, as human beings, use. You may check this page where I try to gather my thinking about AI (sorry this page is written in French for the moment).
An other mysterious word in my research is "learning". Like "intelligence", algorithms do not learn anything in the way human beings or animals learn. In machine learning, learning means computing a set of parameters. Nothing else.
I am working with Prof. F. Pattou and his INSERM Unit 1190 at the University de Lille and CHU de Lille.
Together, we explore how machine learning, reinforcement learning and bandits, may improve health-care, particularly post-surgery patient follow-up.
This collaboration is funded by a set of projects, including this i-site ULNE Bandits for Health (B4H) project. We have designed and set-up a web service to predict weight evolution of obese patients undergoing bariatric surgery. In Sep 2022, Inria published an article on Inria website regarding the Bandits for Health (B4H) project. In 2023, this collaboration will go on much further, funded by the ANR BIP-UP (2023-2026).
In Scool, J. Teigny has worked and T. Soumphonphakdy is working with me as engineers, Patrick Saux is doing his PhD on post-surgery patient follow-up.
I also have had (2019-2022) a collaboration with P. Schegg, PhD student in the Defrost team and the Robocath company, on the control by reinforcement learning of catheters used in surgery. The work on the control of soft robots with reinforcement learning goes on through my collaboration with Defrost.
I have, or had, activities on the following topics related to sustainable development:
In this section, I very briefly describe the topics I worked on in the past.