Designing an effective reward signal for reinforcement learning systems in complex environments is significantly harder than for simple, discrete-outcome games like Go