Incremental Development of Adaptive Behaviors using Trees of Self-Contained Solutions

We present a simple methodology for incremental development of complex adaptive behaviors in robots. The methodology decomposes a given root strategy into a tree of self-contained supporting strategies that can be fully implemented and tested before the next strategy is added. The methodology also uses comparative performance tests for each new strategy relative to its predecessor. The methodology assumes the use of skill modules that can be shared by multiple behavioral layers and produces learning mechanisms that are highly specialized and context dependent. Two example applications of the methodology are presented using simulated robots in the domains of foraging/mapping and conflict resolution. The examples are implemented by hand using a decomposed model of behavior that allows skill modules to be shared while retaining a unique representation of each strategy for excitation and inhibition. Finally we discuss how the solutions produced using this methodology differ from existing behavior-based solutions.