Emergent Hierarchical Control Structures: Learning Reactive / Hierarchical Relationships in Reinforcement Environments

Abstract

The use of externally imposed hierarchical structures to reduce the complexity of learning control is common. However, it is acknowledged that learning the hierarchical structure itself is an important step towards more general (learning of many things as required) and less bounded (learning of a single thing as specified) learning. Presented in this paper is a reinforcement learning algorithm called Nested Q-learning that generates a hierarchical control structure in reinforcement learning domains. The emergent structure combined with learned bottom-up reactive reactions results in a reactive hierarchical control system. Effectively, the learned hierarchy decomposes what would otherwise be a monolithic evaluation function into many smaller evaluation functions that can be recombined without the loss of previously learned information.