Not yet ready for consumption
This is so far a toy project. The goal is to develop a function approximator particularily suitable for representing value functions Q(s,a) (from control and reinforcement learning). Key points are that the approximator is easy to update recursively and to take the argmaxₐ(Q) of.
The approximator consists of a KD-tree, where each cell is a quadratic model. The models are updated recursively once a new datapoint is available. Every now and then, the cell/model with highest covariance (in some sense) is split into two new cells/models.
