Or you can train multiple ones with a kind of evolutionary approach - e.g. you train 100 at the same time and after a day you take the one with the best results, overwrite the rest with it's weights/config/etc. and repeat until satisfied.