Modeling Fake News in Social Networks with Deep Multi-Agent Reinforcement Learning

Code for ICLR2020 submission.

Summmary

The code is found in src/ and is organized into two sub-folders: bias_attack/ and takeover_attack/. Code in each folder is for their respective attack modes. Core elements of this code are:

• env.py: Environment for information aggregation game
• q_agent.py: Q learning agent that implements the recurrent neural network, takes action given observations and updates the network weights to minimize the TD error.
• trainer.py: A training routine that takes in parameters, sets up the Q agent, trains the neural network and periodically saves snapshots of the network. The routine also allows for the post training evaluation of a saved network.

Requirements

Install anaconda.

$ conda create -n new_venv python=3.6
$ source activate new_venv
$ pip install -r requirements.txt

Running code

Adjust parameters in train or test scripts provided in run/ to desired parameters. Copy script into folder in src/ that corresponds to desired attack mode. Run script. Note that scripts provided are examples. Scripts to reproduce results can be obtained by setting parameters appropriately.

License

Code licensed under the Apache License v2.0