If you want say less draws or no advantage for White or less memorization, the easiest solution to me seems to be to just have engines evaluate a few trillion random positions and then filter out those with 0 advantage and low chance of draws in explored lines and then use those as the starting positions.