Immediate reward reinforcement learning for clustering and topology preserving mappings

We extend a reinforcement learning algorithm which has previously been shown to cluster data. Our extension involves creating an underlying latent space with some pre-defined structure which enables us to create a topology preserving mapping. We investigate different forms of the reward function, all of which are created with the intent of merging local and global information, thus avoiding one of the major difficulties with e.g. K-means which is its convergence to local optima depending on the initial values of its parameters. We also show that the method is quite general and can be used with the recently developed method of stochastic weight reinforcement learning [14]