Model-agnostic Fits for Understanding Information Seeking Patterns in Humans
Soumya Chatterjee, Pradeep Shenoy
To appear at AAAI 2021
In decision making tasks under uncertainty, humans display characteristic biases in seeking, integrating, and acting upon information relevant to the task. Here, we reexamine data from previous carefully designed experiments, collected at scale, that measured and catalogued these biases in aggregate form. We design deep learning models that replicate these biases in aggregate, while also capturing individual variation in behavior. A key finding of our work is that paucity of data collected from each individual subject can be overcome by sampling large numbers of subjects from the population, while still capturing individual differences. In addition, we can predict human behavior with high accuracy without making any assumptions about task goals, reward structure, or individual biases, thus providing a model-agnostic fit to human behavior in the task. Such an approach can sidestep potential limitations in modeler-specified inductive biases, and has implications for computational modeling of human cognitive function in general, and of human-AI interfaces in particular.
- PyTorch (Tested with version 1.6.0)
- NumPy (Tested with version 1.91.1)
- Pandas (Tested with version 1.1.4)
- tqdm (Tested with version 4.35.0)
- scipy (Tested with version 1.5.2)
- seaborn (Tested with version 0.8.1)
Download dataset from https://datadryad.org/stash/dataset/doi:10.5061/dryad.nb41c. Create a directory data and place the files TGBE_cardturn_data_54145.mat and TGBE_cardturn_data_40686_robbparam.mat in the data directory.
main.pyhas the main code and generates Figures 1b, 4, 5, 6, 7 in the paper. It has the following arguments
-h, --help show this help message and exit
--approach_avoid Use the data with Approach Avoid Parameters
--tasks TASKS [TASKS ...]
IDs of tasks to be used
--split_frac SPLIT_FRAC
Train-val split fraction
--train_batch_size TRAIN_BATCH_SIZE
Training batch size
--device DEVICE
--lr LR Learning rate
--epochs EPOCHS Number of training epochs
--calc_nll Calculate negative log likelihoods for the model and
the baseline
--all_plots Show all plots. Default: Only show the plots which are
there in the paper
main.py --calc_nllgives the negative log likelihood values for the model and the baseline (Figure 3)main.py --approach_avoidgenerates Figure 9samp_comp_subject_specific_models.pygenerates one data point for Figure 8. Data from multiple runs can be plotted usingsamp_comp_plot.py
Other Files:
cardenv.pycontains our implementation of the game environment which is used for simulationsbaseline_model.pycontains the learnt parameters of the baseline modelget_dataframe.pyis used to preprocess*.matdata files for our codeloss.pycontains the loss functionnetworks.pyhas the models used in the paper for training and simulationsnll_utils.pyis used to calculate negative log likelihoods for Figure 3plot_utils.pyhas all the plotting functionssim_utils.pycontains helper functions for simulating the learnt modelstrain_eval_utils.pyhas helper functions for training and evaluating the modelsutils.pyhas some general purpose helpers
If you find this helpful, please consider citing:
AAAI bibtex coming soon
@misc{chatterjee2020modelagnostic,
title={Model-agnostic Fits for Understanding Information Seeking Patterns in Humans},
author={Soumya Chatterjee and Pradeep Shenoy},
year={2020},
eprint={2012.04858},
archivePrefix={arXiv},
primaryClass={cs.AI}
}