What you describe (sugar + insulin pump) is called a "dual hormone" pump. No such device exists on the open market. So that is yet another challenge. You can a

What you describe (sugar + insulin pump) is called a "dual hormone" pump. No such device exists on the open market. So that is yet another challenge. You can actuate in only one direction: pushing BG lower. PID does actually work for this, but it is outperformed by algorithms that can explicitly consider a patient model (model predictive control, or MPC). I believe closed-loop algorithms using both PID and MPC have been tested on in vivo, and both have been brought to market, although of course all vendors are very tight-lipped about what their algorithm looks like. If you are working in the AP space, you are looking at algorithm performance after an unannounced meal (patient doesn't tell you that he's eaten: how much do you dose?). You will test your algorithm in a virtual patient simulator first. Then, if you are lucky enough to work with the research hospital at your institution, you can test your algorithm on real patients in vivo. The low bar for bringing an AP to the FDA is a provable safety envelope: you will never ever dose too much insulin. Ever. Even if your RL agent learns badly, even if your patient put honey over the CGM site before putting it on. Not having a negative result in the simulator is not enough -- you need a provable bound on dose. This is why the majority of RL algorithms do not work. And why the problem is so hard. You have this super complex system, with a ton of metabolic unknowns. Unknowns over short timescales, like food response, exercise, illness, stress, and so on. And over long timescales, like your diet changes over time, your body changes, etc. And you have huge variation between patients, with different genetics, diet, and astrological sign between people. Boy it would be nice to learn those unknowns with a computer, which is what every RL paper says about the AP. But: if you get it wrong once, it's all over, and further, if you really want to play, you have to prove that you will never get it wrong. That's why every AP system out there now is very conservative, and why none of them are using exotic ML techniques. And why, with the standard classical toolkit, the problem is very difficult. Although it's a bit old, this is a pretty good overview of where the field is: https://diabetesjournals.org/care/article/37/5/1191/38217 https://diabetesjournals.org/care/article/37/5/1191/38217