Behavioral Programming: Enabling a "Middle-Out" Approach to Learning and Intelligent Systems

Abstract Motivated by biology and a study of complex systems, intelligent behavior is typically associated with a hierarchical structure, where the lowest level is usually characterized by continuous-variable dynamics and the highest by a logical decisionmaking mechanism. Consistent with biological findings, we instantiate a particular version of such a hierarchy: a “Middle-out Architecture.” Significant there is the plasticity of circuits on both layers, as well as plasticity in their interactions. Intelligent systems must accomplish two broad tasks within this architecture. Based on experience, they must tune their lower-level, continuous, sensorimotor circuits to produce local behaviors that are viable and robust to slight environmental changes. Predicated on focus of attention on sensory input, they must modulate and coordinate these local behaviors to maintain situational relevance and accomplish global goals. Hence, if we are to understand learning and intelligent systems, we must develop ways of behavioral programming (taking symbolic descriptions of tasks and predictably translating them into dynamic descriptions that can be composed out of lower-level controllers) and co-modulation (simultaneously tuning continuous lower-level and symbolic higher-level circuits and their interactions). Herein, we begin the study of the first of these problems,