Re-usable features in a hierarchical concept network for autonomous learning in complex games

The use of re-usable features to define conceptual elements is a recognised trait in many models of semantic memory, and provides advantages in efficiency of representation, and a manner to preserve links between related concepts. In order to form scalable and generalisable representations, autonomous systems are advantaged by the ability to re-use features, and to develop such a network of features autonomously. Existing learning systems that build knowledge structures in a reinforcement based environment tend to use separately defined rules, rather than re-use of shared features. The system described is a form of Learning Classifier System, based on the Activation-Reinforcement Classifier System that reinforces rules according to separate properties of expected reward and accessibility. This provides a useful platform for examining the construction of rules from re-used features. An implementation is described that constructs a network of features, that are used to define rules. This is able to operate successfully on the game of Dots and Boxes, providing stable operation and the ability to activate rules from a body of 4000 autonomously developed features. Examining the network produced shows a scale-free connectivity distribution, which is a property common in human semantic networks.