Explaining by Imitating: Understanding Decisions by Interpretable Policy Learning

Authors: Alihan Hüyük, Daniel Jarrett, Cem Tekin, Mihaela van der Schaar

ICLR 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Through experiments on both simulated and real-world data for the problem of Alzheimer’s disease diagnosis, we illustrate the potential of our approach as an investigative device for auditing, quantifying, and understanding human decision-making behavior.
Researcher Affiliation Academia Alihan Hüyük University of Cambridge, UK ah2075@cam.ac.uk Daniel Jarrett University of Cambridge, UK daniel.jarrett@maths.cam.ac.uk Cem Tekin Bilkent University, Ankara, Turkey cemtekin@ee.bilkent.edu.tr Mihaela van der Schaar University of Cambridge, UK Cambridge Centre for AI in Medicine, UK The Alan Turing Institute, UK; UCLA, USA mv472@cam.ac.uk
Pseudocode Yes Algorithm 1 Bayesian INTERPOLE
Open Source Code No The paper does not provide an explicit statement about releasing its own code or a link to a repository for the INTERPOLE method. It mentions an "off-the-shelf POMDP solver available at https://www.pomdp.org/code/index.html" used for a benchmark algorithm, but this is not their own implemented code.
Open Datasets Yes For our real-world setting, we consider the diagnostic patterns for 1,737 patients during sequences of 6-monthly visits in the Alzheimer’s Disease Neuroimaging Initiative [48] database (ADNI).
Dataset Splits No The paper describes the datasets used (ADNI, DIAG, BIAS) and their sizes (e.g., "1,737 patients", "1,626 patients", "100 demonstration trajectories", "1000 demonstration trajectories"), but it does not specify explicit percentages or counts for training, validation, or test splits. It mentions aspects of data handling like filtering and MCMC sampling for benchmark algorithms, but not standard dataset splits for reproducibility.
Hardware Specification No The paper does not provide specific details about the hardware used for running the experiments, such as GPU or CPU models, memory, or cloud computing specifications.
Software Dependencies No The paper mentions software components like "Adam optimizer", "LSTM unit", and "MCMC" techniques, and refers to "an off-the-shelf POMDP solver available at https://www.pomdp.org/code/index.html". However, it does not provide specific version numbers for any programming languages, libraries, frameworks, or solvers used, which are necessary for reproducible software dependencies.
Experiment Setup Yes For R-BC: "The network consists of an LSTM unit of size 64 and a fully-connected hidden layer of size 64. We minimize the cross-entropy loss... using Adam optimizer with learning rate 0.001 until convergence, that is when the cross-enropy loss does not improve for 100 consecutive iterations." For DIAG: "η = 10, µa=(s+) = 0.5, and µa (s ) = µa+(s+) = 1.3."