2020 Benchmark Proposal: Artificial Pancreas by Souradeep Dutta
The property is that the network should be monotonically negative for insulin inputs. (eg hold all insulin inputs constant with input range 0-0.1, hold a glucose trace constant (given the below criteria)), and as you increase insulin in one location, the predicted output should decrease. Glucose values are allowed to vary [40,400], with a max difference of 25 between two consecutive inputs eg: G(t) vs G(t-5). These networks all take $7$ glucose inputs G(t), G(t-5), ... , G(t-30) and $7$ insulin inputs u(t), .., u(t-30). They have a single output that represents a blood glucose level prediction G(t+60). Three different networks are provided. The networks use the Sherlock format documented here: https://github.com/souradeep-111/sherlock a) M1: A network with two dense layers. This network is in the file M1_Regular_APNN.nt b) M2: A network with two dense layers with the first layer separated for insulin and glucose inputs. This network is in the file M2_SplitLayer_APNN.nt c) M3: A network with same topology as M2 but has been designed so that the output is guaranteed to be "monotonic" w.r.t to the insulin inputs. This network is in the file M3_WeightCons_APNN.nt
Model files:
https://drive.google.com/open?id=1PilQa-R5zezamaQDMhLVaBzvKAjKRIDG
Paper:
This benchmark is simply a feedforward network, with 3 different types of networks, correct? If I understand correctly, although the filename extension is different, the network format is the same that you used to use (.txt or no file extension vs .nt).
Given this benchmark is primarily a feedforward network without plant, we'd suggest excluding this one for this iteration.