Oberseminar 13.01.2015
14:15h - 15:45h
Raum 057, Oet. 67
-------------------------------------------------------
Stefanie Neubert - Solving Relational Reinforcement Learning Problems with a Combination of Incremental Decision Trees and Generalization
Masterarbeit, betreut von Lenz Belzner
Reinforcement learning is a wide-spread and powerful research area in artificial intelligence. However, the potential of reinforcement learning – and learning techniques in common – is constrained by the scope of the problems. Large problem domains cause its multiple in storage space and learning time. The use of a more powerful relational representation of the problem increases the scope further. Several approaches exist to overcome this problems. The usage of efficient storage structures or clustering of sub problems via generalization are two of them.
The underlying thesis combines the two approaches of efficient storage structures and generalization. It presents a learning algorithm, which uses relational trees to store the classes of the problem. The trees are built incremental and online. The algorithm can deal with a relational problem representation and regards sorts and sort-orders of objects. The decision criteria of the trees are created, generalized and selected by the algorithm, without any help of the user. The precision of the decision criteria can be adapted with parameters.
The required storage space of the algorithm is small even if many different problem states were visited. The generalization mechanism improves the convergence behaviour of the algorithm significantly. The algorithm converges after fewer episodes and the trees contain less split criteria.
