Brief
On Google Brain's ventilator data (inputs u_in, u_out, R, C; target: the pressure at each time step of a breath), recover the discrete pressure grid (pressure only ever takes values from a fixed set) and fit the simulator's PID dynamics so that whole breaths reconstruct exactly. A PID controller sets its output from the current error (P), the accumulated past error (I) and how fast the error is changing (D). Then train a bidirectional LSTM / conv-transformer with auxiliary losses (extra prediction targets that help training) on the breaths that did not match. Submit late to the Kaggle leaderboard.
VICTORY CONDITIONS
- late-submission score inside the top 1% of the private LB; ≥ 1 documented PID-matching insight in a public notebook (Kaggle private-leaderboard MAE (mean absolute error), scored on the inspiratory (breathing-in) phase only; % of test breaths reconstructed exactly).
- A public repository with a README that shows your result and how to reproduce it.
Terms in plain words
- PID controller
- A feedback rule that sets a control signal from the current error, the sum of past errors and the error's rate of change.
- u_in / u_out
- The inspiratory valve opening (how much air is let in) and whether the exhalation valve is open.
- R and C
- Lung resistance and compliance, two settings that describe how the simulated lung responds to air.
- Private leaderboard
- The hidden part of a Kaggle test set that decides final rankings, so you cannot tune against it.
- Late submission
- Scoring an entry on Kaggle after the competition has closed; it shows your rank without counting officially.
Why it rates ★★★☆☆
Reverse-engineering the process that generated the data is an unusual, impressive skill, even though the trick is public.