Evolving neuromodulatory topologies for reinforcement learning-like problems