Efficient Reinforcement Learning using Relational Aggregation
Martijn van Otterlo · University of Twente Research Information · 2003
Much research in Reinforcement Learning (RL) has focused on learning algorithms and generalization using simple representation languages for states and actions. Recently, there is much interest in various kinds of abstraction. Abstractions over time or primitive actions, e.g. in hierarchical RL, are useful ways to abstract over specific sub-actions or time. Currently there is also interest in using more powerful representation languages for abstraction in RL, in which subsets of first-order logic are used for representing sets of states and actions. For an overview of these methods, see (van Otterlo., 2003; van Otterlo, 2002).