This repository includes a customizable simulated environment for testing different aspects of the agent/environment interaction in addressing the behavioral nudging personalization contextual bandit problem. The code was written around 2022 and may be a bit outdated now.
The findings from SXF contributed to the papers
- Examining Policy Entropy of Reinforcement Learning Agents for Personalization Tasks
- Increasing Entropy to Boost Policy Gradient Performance on Personalization Tasks
- On the Unreasonable Efficiency of State Space Clustering in Personalization Tasks
- Install conda / pip requirements via
conda env create -f environment.yml - Activate conda environment with
conda activate synthetic-experimental-framework - Modify experiment configuration in
configs/config.ymlas needed - Run a new experiment via
python -m run_experiment, or load a recorded experimentexp_nameviapython -m run_experiment --load exp_name - Results of the experiment can be found in
exp_data/exp_name/directory
environment.yml--- list of the required packagesconfigs/config.yml--- config file containing the experiment/envirnment/agent parametersrun_experiment.py--- main module to run the experimentexp_data/--- directory containing the experiment data, images, checkpoints
experiment_component/experiment.py--- setup the experimentexperiment_component/data_visualization.py--- compute and report results of an experiment
environment_component/environment.py--- setup the environmentenvironment_component/state_space.py--- setup the state space for the environmentenvironment_component/action_space.py--- setup the action space for the environmentenvironment_component/reward_function.py--- setup the reward function for the environmentenvironment_component/feedback_signal.py--- setup the feedback signal that is given to the agent
agent_component/agent.py--- setup the agentagent_component/network_architecture.py--- setup the policy for the agentagent_component/loss_function.py--- setup the loss function for the agent