POETREE: Interpretable Policy Learning with Adaptive Decision Trees

Authors: Alizée Pace, Alex Chan, Mihaela van der Schaar

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental This policy learning method outperforms the stateof-the-art on real and synthetic medical datasets, both in terms of understanding, quantifying and evaluating observed behaviour as well as in accurately replicating it with potential to improve future decision support systems.
Researcher Affiliation Academia Aliz ee Pace ETH AI Center, Switzerland ETH Z urich, Switzerland MPI for Intelligent Systems, T ubingen, Germany alizee.pace@ai.ethz.ch Alex J. Chan University of Cambridge, UK ajc340@cam.ac.uk Mihaela van der Schaar University of Cambridge, UK Cambridge Centre for AI in Medicine, UK The Alan Turing Institute, UK mv472@cam.ac.uk
Pseudocode Yes Algorithm 1: Tree growth optimisation.
Open Source Code Yes Code is made available at https://github.com/alizeepace/poetree and https:// github.com/vanderschaarlab/mlforhealthlabpub.
Open Datasets Yes First, as ground-truth policies πE are inaccessible in real data, a synthetic dataset of 1000 patient trajectories over 9 timesteps, denoted SYNTH, was simulated. Next, a real medical dataset was explored, following 1,625 patients from the Alzheimer s Disease Neuroimaging Initiative (ADNI), as in H uy uk et al. (2021) for benchmarking. Finally, we also consider a dataset of 4,222 ICU patients over up to 6 timesteps extracted from the third Medical Information Mart for Intensive Care (MIMIC-III) database (Johnson et al., 2016)...
Dataset Splits Yes Model hyperparameters in Table 10 were optimised through grid search on validation datasets random subsets of 10% of the training data.
Hardware Specification Yes Experiments were performed on a Microsoft Azure virtual machine with 6 cores and powered by a Tesla K80 GPU.
Software Dependencies No The paper mentions software like JAX, scikit-learn, and a POMDP solver but does not provide specific version numbers for these dependencies.
Experiment Setup Yes Model hyperparameters in Table 10 were optimised through grid search on validation datasets random subsets of 10% of the training data. We used the Adam optimiser (Kingma & Ba, 2015) with a step size of 0.001 and a batch size of 32.