Comment by algo_trader

Where do they benchmark code candidates? This is hard, noisy and expensive (I see example reward for code size).

I am not an expert, but bayesian learning maybe more appropriate for such an expensive-sampling environment?