Tabular learning methods such as Q-Learning quickly become impractical when faced with large state spaces. Domain information and careful representations can lead to reduction of state space size, but often by sacrificing the generality of the learning method. Ultimately the learning system is designed to complete a certain task (reach a certain goal). Arity based analysis of goals offers a way of significantly optimizing the memory usage of Q-Learning and allows us to approach problems of larger state spaces with our Discretized Q-Learning method.

In this talk we will demonstrate an experimentation system for interaction-learning like tasks, introduce arity based goal analysis, possible alternatives to the reinforcement signal as a way of goal specification, and successful application of DQ-learning to problems with large state spaces, e.g. 5x5 Lights Out puzzle.