Background and activities
Project proposals for ITK 2019-2020
A novel approach to Reinforcement Learning
Reinforcement Learning (RL) is a fast-developing field of research, which deals with exploiting data to improve decision-making for dynamic systems. While most RL approaches are based on using Deep Neural Network to support the decision making, we are developing alternative approaches based on constrained optimization. Our ambition with these new approaches is to be able to impose safety constraints for AI-based decision making. A number of questions are, however, still open. One open question we will start investigating in this project is the developments of rich parametrizations of the constrained optimization problem, and their effect on the performance of the AI agent.
The project is suitable for two students working in a team. The student(s) taking part in this project will discover cutting-edge computer tools for optimization and optimal control, and learn some of the theory behind RL and optimal control.
NTNU ITK supervisor: Sebastien Gros (Sebastien.gros@ntnu.no)
Risk-management in predictive control
In the context of predictive control for mission planning (e.g. autonomous vehicle, drone, etc.), the risk associated to a selected plan can be of crucial importance for the operator. A very useful view of the problem is to measure the risk in terms of the probability of failing the mission, and the cost of failure. Assessing accurately the probability of failing a mission under a given plan (in addition to e.g. some low-level control systems) is, unfortunately, fairly difficult to do at a reasonable computational cost. In this project, we will explore novel ideas to perform this estimation at a lower computational cost.
The project is suitable for two students working in a team. The student(s) taking part in this project will learn concepts from advanced statistics, and Markov Processes, and learn how to handle these problems in the computer.
NTNU ITK supervisor: Sebastien Gros (Sebastien.gros@ntnu.no)