a2rl 1.0.1

Make recommendations for sequential decision problems using offline data

Amazon Accessible RL SDK

[ Documentation | PyPI ]

Amazon Accessible RL (A2RL) is an open-source Python package for sequential decision making problem using offline time-series data. It focuses on offline RL using state-of-the-art generative transformer technology – the same technology behind GATO, trajectory transformer and decision transformer.

A2RL guides you through problem formulation, conduct initial data analysis to see if a solution is possible, use the data to train a simulator (aka digital twin) based on the data, and providing recommended actions.

Installation

pip install a2rl

Usage

You should start by formulating your problem into states, actions, and rewards (see the online documentation). Then, prepare a dataset that reflects the formulation, using A2RL's Pandas-like API.

A synthetic dataset is included to help you quickly jump into the end-to-end workflow:

import a2rl as wi
from a2rl.utils import plot_information

# Load a sample dataset which contains historical states, actions, and rewards.
wi_df = wi.read_csv_dataset(wi.sample_dataset_path("chiller")).trim().add_value()
wi_df = wi_df.iloc[:1000]  # Reduce data size for demo purpose

# Checks and analysis
plot_information(wi_df)

# Train a simulator
tokenizer = wi.AutoTokenizer(wi_df, block_size_row=2)
builder = wi.GPTBuilder(tokenizer, model_dir="my-model", )
model = builder.fit()
simulator = wi.Simulator(tokenizer, model, max_steps=100, reset_coldstart=2)

# Get recommended actions given an input context (s,a,r,v,...s).
# Context must end with states, and its members must be tokenized.
custom_context = simulator.tokenizer.df_tokenized.sequence[:7]
recommendation_df = simulator.sample(custom_context, 3)

# Show recommendations (i.e., trajectory)
recommendation_df

Help and Support

• Contributing
• Apache-2.0 License