Learning predictive partitions for continuous feature spaces
Björn Weghenkel, Laurenz Wiskott · The European Symposium on Artificial Neural Networks · 2014
Any non-trivial agent (biological or algorithmical) that in- teracts with its environment needs some representation about its current state. Such a state should enable it to make informed decisions that lead to some desired outcome in the future. In practice, many learning algo- rithms assume states to come from a discrete set while real-world learning problems often are continuous in nature. We propose an unsupervised learning algorithm that finds discrete partitions of a continuous feature space that are predictive with respect to the future. More precisely, the learned partitions induce a Markov chain on the data with high mutual information between the current state and the next state. Such predictive partitions can serve as an alternative to classical discretization algorithms in cases where the predictable time-structure of the data is of importance.