2020 | CoLeSIAw : Continuously Learning Complex Tasks via Symbolic Analysis

Axe : SciLex, IID
Sujet : IID-1 Machine learning, Scilex-2 Continuous versus discrete systems

Directeurs de thèse : Sergio Mover
Laboratoire gestionnaire : LIX
Autre partenaire : U2IS
Doctorant :
Début : 2020
Productions scientifiques :
Ressources :

Contexte : Fully autonomous robots have the potential to impact real-life applications, like assisting elderly people. Autonomous robots must deal with uncertain and continuously changing environments, where it is not possible to program the robot tasks. Instead, the robot must continuously learn new tasks. The robot should further learn how to perform more complex tasks combining simpler ones (i.e., a task hierarchy). This problem is called lifelong learning of hierarchical tasks.

The existing learning algorithm for hierarchical tasks are limited in that: a) they require the robot to execute a large number of real actions to sample the continuous state space of observations, hence requiring a lot of time; b) they cannot deal with subspaces without continuous interpolation, as it is the case for a hierarchy of tasks.

In this Ph.D. project we will exploit the intuition that set-based reasoning of the continuous space can reduce the number of samples required to learn a hierarchy of tasks and allow for a more effective planning of the robot tasks, further handling discontinuities in the task hierarchies. We will design new algorithms to effectively explore the task hierarchies and new reachability algorithms for data-oriented models such as neural networks.