Organization responsible: TNO

People involved:

• Frans Groen (promotor)
• Leon Kester (co-promotor
• Bram Bakker (researcher)
• Harm van Seijen (PhD student)
• Shimon Whiteson

Project description:

Our general research approach is guided by the belief that the ultimate goal of AI research should not be to build the most intelligent system possible, but to solve real, complex problems as efficiently as possible. This means on the one hand a search for efficent algorithms, and on the other hand a search for algorithms that make optimal usage of all available (domain-specific) prior knowledge.

In the previous year we have developed a reinforcement learning architecture in which the agent can dynamically switch between qualitatively different state representations.
There are two dimensions in which we want to evolve this research. The first dimension is the problem complexity. So far, we are dealing only with rather simple toy problems. We want to evolve this slowly into the direction of real applications, by gradually increasing the complexity of the problems we are investigating. This means we will focus more on sub-optimal techniques that are able to find a good approximate solution within a reasonable amount of time rather than optimal solutions.

Besides investigating sub-optimal techniques, we will pay specific attention to the optimal usage of prior knowledge. In our experience often not all available prior knowledge is used. This unused prior knowledge will serve as a starting point for searching for new, more efficient algorithms.

The second dimesion we want to make progress on is theoretic grounding of our approach. We want to ground our work more strongly in a theoretic framework. This will aid our fundamental understanding of the switching and will help us convince others to use our switching method. We foresee the need to cooperate with other researchers to make substantial progress in this dimension.

Publications: