I have a model (which I would initialize with "best guess" parameters) which produces a real-valued output (which I need to be real-valued - I can't switch to a different representation), but for which I can only obtain binary (good / bad) feedback from human evaluation. It's very easy for a human to tell if the output of a model is good or bad for a given example, but difficult for them to put an exact number on how good / bad it is. I can't change the output of the model, but I CAN freely choose which model to use / the architecture of the model.
How can I design the model so that it can learn from that feedback? I see 3 papers that utilize binary feedback to train classifiers, but I have not seen any which use the same to train on real-valued output.