minregret / tigercontrol Goto Github PK

View Code? Open in Web Editor NEW

38.0 9.0 3.0 19.1 MB

Google AI Princeton control framework

Home Page: https://tigercontrol.readthedocs.io/en/latest/

License: Apache License 2.0

Python 16.30% Jupyter Notebook 83.70%

control benchmarking pybullet autograd machine-learning python jax

tigercontrol's Introduction

TigerControl

NOTE

This library is deprecated. Please refer to our new control library, deluca.

TigerControl is an open-source framework for benchmarking control algorithms in simulated and real settings, and is available for anyone to download and use. By reducing algorithms to a set of standard APIs, TigerControl allows the user to quickly switch between controllers and tasks while running experiments and plotting results on the go, and for quick and simple comparison between controller performances. TigerControl also comes with built-in standard control algorithms for comparison or other use.

Overview

Although there are several machine learning platforms that aid with the implementation of algorithms, there are far fewer readily available tools for benchmarks and comparisons. The main implementation frameworks (eg. Keras, PyTorch) provide certain small-scale tests, but these are in general heavily biased towards the batch setting. Other platforms such as OpenAI Gym are great for the testing of reinforcement learning controllers, but as many online algorithms depend on access to interal dynamics or loss derivatives, this too is suboptimal. TigerControl exists to fill this gap — we provide a variety of evaluation settings to test online control algorithms on, with more diversity and generality than any other online benchmarking platform.

Installation

Quick install (excluding PyBullet)

Clone the directory and use pip or pip3 to set up a minimal installation of TigerControl, which excludes PyBullet control environments.

    git clone https://github.com/MinRegret/tigercontrol.git
    pip install -e tigercontrol

You can now use TigerControl in your Python code by calling import tigercontrol in your files.

Full install (including PyBullet)

Before installing TigerControl's dependencies, create a new conda environment in order to avoid package conflicts.

    conda create -n name-of-your-environment python=3.x
    conda activate name-of-your-environment

Next, clone the GitHub repository and install the full package.

    git clone https://github.com/MinRegret/tigercontrol.git
    pip install -e 'tigercontrol[all]'

For more information

To learn more about TigerControl and how to incorporate it into your research, check out the Quickstart guide in the tigercontrol/tutorials folder. Alternatively, check out our readthedocs page for more documentation and APIs.

tigercontrol's People

Contributors

Stargazers

Watchers

Forkers

paula-gradu vvanirudh davtalab

tigercontrol's Issues

which gpc version to use?

Hi,

I'd like to try out Algorithm 1 in https://arxiv.org/pdf/1902.08721.pdf.

I see two files:
https://github.com/MinRegret/TigerControl/blob/master/tigercontrol/controllers/gpc.py
https://github.com/MinRegret/TigerControl/blob/master/tigercontrol/controllers/gpc_v2.py

Which version is recommended?

Thanks!

finish quadcopter

LDS EM algorithm model

Fix notebooks (split into smaller ones)

Scoreboard for basic controllers (LQR/GPC)

Experiment plot method

Implement a plot method for Experiment class, which allows users to immediately plot the results of their experiments.

Migrate to google/experimental

Adding nearest-neighbor to problem and model registry's in case of misspelled IDs

Crypto dataset Problem

Add download utils and Problem class for cryptocurrency dataset (mostly bitcoin)

Physics Engine

LONG TERM FEATURE - Create physics engine for control simulation (ex MuJoCo/Unity) to allow for future control Problems

Model PredictZero and LastValue unit tests

Add unit tests to new Model classes

Adaptive disturbances model

Problem: Add 2d drone (planar quadrotor)

Simple control Problem

Find a simple control Problem and implement it (simple enough that it can be used as a debugging tool)

tigercontrol.error.UnregisteredObject: No registered <TigerControl Problem Registry> with ID: CartPoleSwingup-v0, did you mean PyBullet-CartPoleSwingup-v0?

Hi,

I'm getting the next error when running /Users/rgap/Desktop/LIBS/tigercontrol/tigercontrol/models/control/tests/test_mppi.py

tigercontrol.error.UnregisteredObject: No registered with ID: CartPoleSwingup-v0, did you mean PyBullet-CartPoleSwingup-v0?

Preserve user input-output shape

Method: Make LDS a master class with inherited subproblems

Kalman Filter Model

Implement a Linear Quadratic Regulator Model for control tasks, see https://en.wikipedia.org/wiki/Linear%E2%80%93quadratic_regulator

make system identification compatible + API

Feature: Infinite time horizon LQR

tutorials/demos

notes:

our experiments (using tigercontrol)
nataly's experiments

Model subclasses

The update method varies for different types of models. To resolve this issue, we have made the design choice of implementing model subclasses (TimeSeries, Control, etc) each with the same type of update rule (ex. TimeSeries takes a true label, prediction, params, and loss). Someone should implement these subclasses and place them in a new directory under models/subclasses or some other fitting name.

Bugfix: Experiment inputs not generic

Experiment inputs aren't generic enough to work on arbitrary problems, currently throws errors on several problems, for example LDS because it does not pass it an input at every step.

make kalman filter compatible + API

Method: Fix MPPI

move iLQR to planners

Google documents Design Doc

Autoregressive Model

The Linear model currently has no implementation, and will throw a NotImplementedError every time it is initialized. TODO: Implement a basic online ARMA model and make sure it passes the unit test.

Experiment: Finish adding features/debugging - including making sure it works for Control problems and methods

fix control boosting

Private models

Allow for the creation and registration of custom models

Custom Model implementation and registration

Add feature that allows users to implement their own Models and register it automatically to CTSB (locally) - John

ReadTheDocs (Jax Documentation)

Implement experiment and scoreboard classes

Replace asserts with raise Errors

In non-test files (ex Models and Problems) where input is checked for correct shape and type, we have previously been using "asserts" instead of "try-except raise Error". We need to go through all our legacy code and fix this!

Collab tutorial for stock prediction using AutoRegressor

Implement Boosting

tasks

notes:

create folder in directory

Experiment auto-initialize

Automatic smart initialization of Problems and Models for the Experiment class (based on dimensionality of given problem) so that user doesn't have to call problem/model.initialize() for all the inputs

Core: Add to Python package index

Private datasets

New Problem type — allow users to create a CustomProblem instance (name pending) which takes a locally stored dataset, owned by user, and creates a time-series out of it. Example implementation approach could be to let the problem.initialize method take the directory path to the dataset