Implementing Adaptive Capabilities on Agents that Act in a Dynamic Environment

Acting in a dynamic environment is a complex task that requires several issues to be investigated. In this paper, a lifecycle for implementing adaptive capabilities on intelligent agents is proposed, which integrates planning and learning within a hierarchical framework. The integration between planning and learning is achieved by an agent architecture explicitly designed for supporting abstraction. Planning is performed by adopting a hierarchical interleaved planning and execution approach. Learning is performed by exploiting a chunking technique on successful plans. A suitable feedforward neural network selects relevant chunks used to identify new abstract operators. Due to the dependency between abstract operators and already-solved planning problems, each agent is able to develop its own abstract layer, thus achieving an individual adaptation to the given environment.