NclustRL is a toolkit that implements some functionalities to help train agents for n-clustering tasks. It works with Ray's RLlib to train DRL agents.

Ray is a general-purpose framework for distributed computing that implements a known library for hyperparameter tunning, Tune. Furthermore, it implements RLlib, a DRL framework that supports distributed computing and great customization.

NclustRL implements a trainer API for n-clustering that handles all training tasks for the user; a set of default models and metrics; and other helpful functions. Likewise, it provides a set of default configurations for n-clustering tasks available in "configs".

The trainer API aims to provide a simple way of training and testing DRL agents for n-clustering tasks. This class handles all of RLlib's logic and expose only user-friendly methods.

After initialized, the trainer exposes four primary methods:

• Train: Exposes the primary training function. It receives the training parameters that should be passed on to Tune, initiates the training process, manages multiple samples of the same trial, and parses results returning the best performance obtained;
• Load: Imports an agent from a checkpoint for testing;
• Test: Evaluates the accuracy and mean reward and returns the mean and standard deviation for each of these metrics across n episodes.
• Test Dataset: Evaluates the performance in the same way as Test but receives as input a specific dataset from where episodes should be sampled.

This tool can be installed from PyPI:

pip install nclustRL

Here are the basics, for more information check the Experiments available on "Exp".

## Train basic agent

from nclustRL.trainer import Trainer
from nclustRL.configs.default_configs import PPO_PBT, DEFAULT_CONFIG
from ray.rllib.agents.ppo import PPOTrainer
from nclustenv.configs import biclustering

# Inicialize Trainer

config = DEFAULT_CONFIG.copy()
config['env_config'] = biclustering.binary.basic_v2

trainer = Trainer(
    trainer=PPOTrainer,
    env='BiclusterEnv-v0',
    save_dir='nclustRL/Exp/test',
    name='test',
    config=config
)

## Tune agent

best_checkpoint = trainer.train(
    num_samples=8, 
    scheduler=PPO_PBT,
    stop_iters=500,
)

By default this tool implements a model for hybrid proximal policy optimization algorithm, available in "models". This model can be customized, or other models might be implemented and passed in the configs.

GPLv3