Skip to ContentSkip to Navigation
Research Bernoulli Institute Research

Learning in Computational Intelligence

Key words: Machine learning, reinforcement learning, image recognition, robotics
Themes: e-science, e-humanities, healthy aging, energy

Short description:

The field of Machine Learning is currently going through a phase of rapid developments. Many new techniques are emerging, also from our group. A disadvantage of many learning schemes is that they require detailed and time consuming feedback and supervision by a sentient agent, usually a human, who provides “ground truth” for correct actions. There exists a more convenient method of training, borrowed from learning in biological systems, i.e. “reinforcement learning”. In this type of learning, the feedback is simple and coarse, just requiring a reward or punishment. The learning agents have to find out themselves how to adapt their behavior, like animals foraging in their ecological environment. Such agents use a decision policy to select actions based on inputs. Their goal is to optimize the rewards they get through their reward function, which maps changes of the environmental state to scalar reward signals.

In our group we have developed a novel multi agent reinforcement learning system for optimizing traffic lights. This system has been shown to outperform current traffic light control methods by far, and can adapt themselves to changing traffic light patterns. We have also developed learning systems for many games, such as Ms. Pac-Man, Star Craft, Backgammon, and Othello, and the resulting game playing programs demonstrate a very high level of playing strength.

Next to reinforcement learning, we have researched many supervised learning algorithms. Although these methods require human provided labels, in many cases datasets are available or crowd sourcing can be used to obtain those labels. We have developed an image recognition system for classifying different objects and scenes from the Corel image data base. So far our system has obtained the lowest error rate on this famous image data set. We have also made a facial expression recognition system, classifying emotions from pictures of people. Our system can correctly classify pictures as having one of seven emotions in 95% of the cases.

Group leader: Marco Wiering
Participating researchers: 4
Research Institute: Bernoulli Institute
Faculty: Faculty of Science and Engineering
Graduate school: Graduate School of Science and Engineering (GSSE)

Return to overview of research themes
Go to the next Alice research theme

Last modified:10 January 2020 11.12 a.m.