Architectural Mechanisms for the Dynamic Modification of Action Selection Strategies

Action selection is typically a “built-in” feature of agent architectures and hence not subject to change in instantiated architectures. While often (especially in simple agents) this lack of flexibility does not matter, there are circumstances where dynamic adjustments of an action selection strategy are desirable. In this paper, we present an architecture framework, called APOC, that provides architectural mechanisms for dynamic adjustments of action selection schemes. We show how common action selection schemes (such as “winner-takes-it-all” or hierarchical schemes as used in subsumption architectures) can be implemented in APOC and, furthermore, how they can be modified dynamically. Finally, preliminary results from multi-agent simulation studies of such architectures are presented to demonstrate the utility of dynamically modifiable action selection schemes.