Encoding Human Behavior in Information Design through Deep Learning
Authors: Guanghui Yu, Wei Tang, Saumik Narayanan, Chien-Ju Ho
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Through extensive simulation, we show that HAIDNet can not only recover information policies that are near-optimal compared with known analytical solutions, but also can extend to designing information policies for settings that are computationally challenging (e.g., when there are multiple receivers) or for settings where there are no known solutions in general (e.g., when the receiver behavior does not follow the Bayesian rational assumption). We also conduct real-world human-subject experiments and demonstrate that our framework can capture human behavior from data and lead to more effective information policy for real-world human receivers. |
| Researcher Affiliation | Academia | Guanghui Yu Washington University in St. Louis guanghuiyu@wustl.edu Wei Tang Columbia University wt2359@columbia.edu Saumik Narayanan Washington University in St. Louis saumik@wustl.edu Chien-Ju Ho Washington University in St. Louis chienju.ho@wustl.edu |
| Pseudocode | No | The paper describes the HAIDNet framework and optimization procedures in detail but does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | The specification of hyperparameters and implementation details are included in the appendix. We have also included the source code in the supplementary materials. |
| Open Datasets | No | The paper describes generating '102,400 uniformly drawn problem instances' for training and collecting data from human-subject experiments, but it does not provide concrete access information (like a link, DOI, or formal citation with author/year) for these datasets, nor does it refer to a well-known public dataset. |
| Dataset Splits | Yes | The model is trained on 100 batches of size 1024, for a total of 102, 400 uniformly drawn problem instances (i.e., data points for training). Evaluation of the model is conducted on a test set consisting of 1000 problem instances. The hyperparameters are tuned by using the validation dataset. |
| Hardware Specification | Yes | The reported numbers are performed on the machines with Intel(R) Xeon(R) Gold 6148 CPU (2.40GHz) and a Tesla V100-SXM2-32GB GPU. |
| Software Dependencies | No | The paper mentions using 'Re LU activation functions' and the 'Adam optimizer' and 'neural network', which are common software components, but it does not specify any version numbers for these or other software libraries (e.g., Python, PyTorch, TensorFlow versions). |
| Experiment Setup | Yes | To optimize HAIDNet, we train a neural network with 3 fully connected layers employing Re LU activation functions and the Adam optimizer. The model is trained on 100 batches of size 1024, for a total of 102, 400 uniformly drawn problem instances (i.e., data points for training). Evaluation of the model is conducted on a test set consisting of 1000 problem instances. The specification of hyperparameters and implementation details are included in the appendix. The number of nodes for each hidden layers is tuned in the range of {64, 128, 256, 512, 1024}, and the initial learning rate is tuned in the range of {0.001, 0.002, 0.005, 0.01, 0.02, 0.05, 0.1}. When the human descriptor is Bayesian rational, we use softmax to smoothen the argmax operator. ...we increase β gradually from 10 to 1000 exponentially in first 100 epochs of training, that is βi = 101+ i 50 in i th epoch, and maintain β = 1000 for remaining training process. |