jundev1l2l / npucb

Neural Processes Upper Confidence Bound Algorithm for Contextual Bandits (NP-UCB)

Pytorch Implementation of UCB Algorithm with Neural Processes (NP-UCB) for Contextual Bandits

This repository contains the codes for my master thesis paper, "Empirical Study on Contextual Bandit Algorithms with Neural Processes". All the works are done when I am a graduate student of Graduate School of Artificial Intelligence (AIGS), Ulsan National Institue of Science and Technology (UNIST), and a member of Learning Intelligence Machine Laboratory (LIM Lab).

ContextualBandits/ contains the wheel bandit experiment, while R6NewsRecommendation/ contains news recommendation experiment. Each experiment needs two configuration files. One of them sets the default configuration for a group of experiments, while the other overrides the default configuration to apply different conditions for each experiments in a same group. script/train/example/ and script/eval/example/ contains example experiments for training and evaluation, respectively. script/official contains all the experiments reported in my thesis.

In detail, script/official/{experiment_group}/ contains default.yaml and exp_config/{experiment_name}.yaml.

At ContextualBandits/ or R6NewsRecommendation/, run experiment with

python main.py \
--project {project_name_for_logging_wandb} \
--task {train_or_eval} \
--default_config script/official/{experiment_group}/default.yaml \
--exp_config script/official/{experiment_group}/exp_config/{experiment_name}.yaml \
{--any_other_args...}

If you want to override some of the configurations when running the code, give them as arguments.