Learning feature hierarchies under reinforcement

Learning feature hierarchies, where larger features are composed of smaller re-used features, is an important area of study in object recognition and classification, and relates to processes in the human visual system. Established techniques are able to build deep hierarchies using neural networks, such as deep learning based on Restricted Boltzmann Machines, however approaches using other machine learning techniques involving reinforcement are not well established. An approach is presented that uses a form of Learning Classifier System to build a hierarchical feature network, for classification of images using the MNIST dataset. Larger scale representations of rules are composed of re-used smaller elements, in a network of 4,000 features and 2,000 rules. The feature network is developed autonomously, according to reinforcement of rules the features participate in. An implementation is shown using the ARCS classifier system to perform classification of images, using rules based on image templates. A second implementation uses rules with image templates constructed from a hierarchical feature network. This shows effective classification performance, but not as accurate as the best neural network and kernel methods. The implementation shows the ability to construct a hierarchical feature network under reinforcement, and its application to develop a rule population used by a Learning Classifier System. An alternative method for modifying existing rules is shown to substitute for standard mutation and crossover processes, to allow exploration of the rule space more closely related to gradient descent and cognitively related processes, rather than the genetic analogy commonly used in learning classifier systems.