Organization responsible: UM, UvT

People involved:

• Jaap van den Herik (promotor, project coördinator)
• Ben Torben-Nielsen (PhD student)
• Marc Ponsen (PhD student)
• Nyree Lemmens (PhD student)
• Katja Verbeeck
• Eric Postma

Project description:

• Adaptivity and computational complexity in neurons. Adaptive behaviour in robots is often implemented by means of Artificial Neural Networks (ANNs). These ANNs consist of very simple and highly abstract computational building blocks –the neurons- and connections between them. In order to increase the adaptivity of such a system, specific network topologies in combination with mathematical learning rules are used. The basic building block –the neuron- is highly abstract but crucial. Inspiration for this methodology comes from the brain. However, in the brain the computational building blocks themselves are already capable of performing very complex and adaptive behaviour. In this research track we aim at identifying principles underlying the complex behaviour in real single neurons, and applying these principles in an artificial systems.
• Relational reinforcement learning. Relational learning methods are a promising new approach to machine learning. Its capability to capture more naturally relations between objects in a domain, suggest this is a suitable learning paradigm for MAS. We are investigating the benefit of relational learning methods in MAS. Current work shows promising results for relational reinforcement learning in cooperative, fully observable MAS scenarios. To explore the full potential of relational learning in MAS, we are interested in applying it in more challenging domains that offer huge complexity, competitive agents and imperfect information and report on the results.
• Bio-inspired reinforcement learning. Bio-inspired multi-agent behaviour (also known as Swarm intelligence) is a form of reinforcement learning which ensures learning in MAS that is robust and efficient due to the ages of evolution on such behaviour. Moreover, such behaviour is based on low resource usage and self-organisation. Individual agents each perform, in essence, simple behaviour which pose a (not necessarily optimal) solution to the specific problem. However by combining each agent's local actions, near-optimal solutions emerge to complex problems, such as the Traveling Salesman Problem. Furthermore, the solutions emerge within acceptable time limits. Ant Colony Optimization is to date the best-known and most examing bio-inspired system. Although this system has shown its value, there are other bio-inspired systems which could pose an even better solution within the same time-span. We are investigating bee-inspired reinforcement learning and are interested in applying it to combinatorial optimization problems and real-time MAS scenarios which are highly dynamic and have a large search-space.