Kristof Van Moffaert

Kristof Van Moffaert is a former member of the lab who is no longer affiliated with the VUB AI Lab.
Postal address:

Vrije Universiteit Brussel
Artificial Intelligence Lab
Pleinlaan 2
Building G, 10th floor
1050 Brussels

Room: 10G711

My research is focusing on the development of reinforcement learning algorithms for multi-objective problems. In real-life, optimization problems often tend to consist of multiple conflicting objectives: an increase in the performance of one objective implies a decrease in the performance of another objective and vice versa. An example can be found in the area of stock markets, where the manager wants to maximize the returns while at the same time reduce the risks. It is clear that a risky policy has the potential of a high return of investment while at the same time a high probability in money losses as well. In these situations, traditional reinforcement learning approaches often lead to an oversimplification of the problem which on its turn results in unrealistic and suboptimal decisions. In my work, I proposed a series of algorithms that allow to discover one or more trade-off solutions that balance the objectives in the environment. We cluster the algorithms in two categories, i.e., single-policy and multi-policy, based on the fact whether the algorithms obtain a single compromise or a set of solutions, respectively. 

Currently, I am affiliated with the PERPETUAL project. This project aims project aims at developing methods and algorithms which allow a wide range of systems around you to optimize their balance between delivered functionality and resource utilization based on automatic usage‐profiling. This trade-off is learned in accordance to the presence of the user and therefore the exploration should be conducted carefully. This is called morphing and consists of reducing the selection probability of actions that might lead to a significant increase in the inconvenience level of the user, while exploring.
At the same time, I also work on the SCANERGY project, an FP7 European project between the Vrije Universiteit Brussel and Sensing and Control S.L. (S&C). S&C is a company located in Barcelona, Spain, that is a specialist in home automation, cloud systems, smart grids and machine-to-machine commuciation. In 2014, I worked at the S&C site for 3 months while developing and simulating machine learning techniques that balance the supply and demand of households in the smart grid, the electrical infrastructure of the future. 
In accordance to the previous years, I am the teaching assistent of the Computational Game Theory (Learning Dynamics) course in 2015.


  • We received the best demonstration award for our work on 'SCANERGY: a Scalable and Modular System for Energy Trading Between Prosumers' at the Autonomous Agents and Multi-Agent Systems conference (AAMAS) 2015
  • The Journal of Machine Learning Research (Impact factor of 3.420 in 2012) accepted our paper 'Multi-Objective Reinforcement Learning using Sets of Pareto Dominating Policies'.
  • Our paper "Adaptive Objective Selection for Correlated Objectives in Multi-Objective Reinforcement Learning" was accepted at 13th International Conference on Autonomous Agents and Multiagent Systems (AAMAS 2014).
  • Our paper "Scalarized Multi-Objective Reinforcement Learning: Novel Design Techniques" is nominated for the Best Paper Award at IEEE SSCI-2013.

PhD thesis

Multi-Criteria Reinforcement Learning for Sequential Decision Making Problems