Many systems are difficult to control by conventional means because of the complexity of the very fabric of their being. Some systems perform very well under some conditions and then burst into wild, maybe even chaotic, oscillations for no apparent reason. Such systems exist in bioreactors, electro-plating and other application domains. In these cases a model may not exist that can be trusted to accurately replicate the dynamics of the real-world system. BOXES is a well known methodology that learns to perform control maneuvers for dynamic systems with only cursory a priori knowledge of the mathematics of the system model. A limiting factor in the BOXES algorithm has always been the assignment of appropriate boundaries to subdivide each state variable into regions. In addition to suggesting a method of alleviating this weakness, the paper shows that the accumulated statistical data in near neighboring states may be a powerful agent in accelerating learning, and may eventually provide a possible evolution to self-organization.