Tuning reinforcement learning reward signals involves an iterative trial-and-error process conducted by human designers at a higher level than an agent's individual learning