KB Refinement through Active Experimentation