Improving Offline RL by Blending Heuristics