Imitation learning in multi-agent environments
Joosep is originally from Estonia and is currently studying a PhD under the Institute for Digital Technologies. Prior to his studies at Loughborough, he was studying Computer Engineering at Tartu University.
Joosep holds a Bachelor of Science in Engineering from University of Tartu. He did his thesis on automatic speech-based emotion recognition, which sparked his interest in Artificial Intelligence and it's applications. During his studies he also managed to work as a developer, gaining valuable experience in software development and system administration.
PhD research description
A fun way to spend your afternoon is to try and teach swear words to a parrot.
While the parrot has little understanding of what you are saying, it nevertheless attempts to mimic the sound of the words it hears.
Imitation learning is a branch of machine learning where the main learning objective is to learn from a demonstrator or demonstrators in order to perform a specific task, be it folding laundry, playing video games or driving cars.
A common theme among the tasks is that the way to perform them is hard to describe formally. Instead of trying to figure out a precise way of describing the solutions as algorithms or programs, imitation learning intends to learn the correct behaviour from looking at demonstrations alone.
While the idea is simple enough to understand intuitively, imitation learning has many challenges. For example, consider driverless vehicles. Most, if not all, of the data collected for driverless vehicles represents 'proper driving'. This means that it is impossible for the machine learning algorithm to recover from disastrous situations it has never seen before. Some algorithms address this problem by having an expert demonstrator correct the mistakes of the learner, but this introduces its own (if the expert is a human and the learner makes a lot of mistakes, then it will query the expert very often).
Awards, grants or scholarships received
Joosep was awarded a Loughborough University studentship in July 2018.
Papers, publications and articles
Joosep Hook, Fatemeh Noroozi, Onsen Toygar, and Gholamreza Anbarjafari (in press). Automatic speech based emotion recognition using paralinguistics features. Bulletin of the PolishAcademy of Sciences Technical Sciences, 2018
Interests and activities
Joosep is interested in anything remotely or closely related to computers and using them to create useful tools for the people that use them.