Apply blurring over the whole context space to try and prevent oscillations between very similar states Use scoring for picking context directions rather than lowest danger weighting