v.1.3.3

MLflow is part of ReinforceUI-Studio (Finally)

• MLflow Integration: Added MLflow support with toggle controls, auto server launch, and dashboard access directly from the UI.
• Training Enhancements: Training config updated to support MLflow logging with consistent algorithm naming.
• Documentation Updates: README improved with MLflow examples, cleaner layout, and streamlined installation instructions.
• We have new badges in the README

Changes:

• Integration of Mlflow by @dvalenciar in #38

Full Changelog: v1.3.2...v.1.3.3

• pip package: version 1.3.3

v1.3.2

Reinforce UI-Studio as a Python package

• Update directories, links and paths to match the logic of PyPI
• Create a build the Python package and make it public here
• Update internal paths
• Update readme

Full Changelog: v1.3.1...v1.3.2

Multi-Algorithm Training

What's New:

• Added the capability to handle training of several algorithms simultaneously.
• Saved a global session_config.json containing shared training parameters.
• Created individual folders per algorithm with their own config.json and saved models/logs separately.
• Automatically generate a final comparison plot after training completes if multiple algorithms were run

Full Changelog: v1.3.0...v1.3.1

New UI style

What's New:

• The graphical interface style has changed significantly.
• New, clean, and fresh look.
• White background colour.
• Consistent style and shapes.
• Blue is the primary colour throughout the project.
• Introduction to the main window, which is now very cool.

Full Changelog: v1.2.2...v1.3.0

1.2.2

v1.2.2

• Fix bug on loops
• Improve code in GUI
• Reduce redudant code in GUI
• Add examples of results
• include Flake8 rules and check

Full Changelog: v1.2.1...v1.2.2

Discrete Action Space & DQN Support

Summary

This release introduces support for discrete action spaces in Gymnasium environments and adds DQN as a supported algorithm. It also includes several improvements and bug fixes to enhance stability and usability.

What's New?

New Features

• Added support for discrete action spaces in Gymnasium environments.
• Integrated DQN as part of the list of supported reinforcement learning algorithms.

Improvements & Fixes

• Performance Optimization: Removed unnecessary np.asarray calls when converting data to PyTorch tensors.

• Bug Fix: Fixed an issue in the environment affecting the evaluation loop.

• Action Space Handling :
  -Introduced RescaleAction, ensuring the action space is [-1, 1].
  -Eliminates the need for manual normalization/denormalization inside loops.

• GUI Update:
  - Users now select the algorithm first, followed by the platform and environment for better workflow.

PPO

What's Changed

• Added Proximal Policy Optimization (PPO) to the algorithm list.
• Updated memory buffer to support PPO logic with a flush mechanism.
• Integrated PPO into the main training loop.
• Removed redundant unsqueeze operations in tensor conversion.
• Simplified GUI for PPO, removing unnecessary parameters (G, exploration, b
  
  atch size).
• Refactored training loop, reducing duplicate code for better maintainability.

Full Changelog: PR

v1.1.0

What's Changed

1. Major:

• Added capability to load a directory from previous training and test the policy while rendering the environment.
• Added a new GUI window to load the pre-trained folder.
• Added "human" render capability for the environment.
• A test policy loop from the model was created directly.

2. Other:

• Simplified and refactored code for better readability and maintainability.
• Changed background colour and font colour.

3. Doc

• Add video of new capability of load pre-train model

First Stable Release

First Stable Release (v1.0.0)

• ✅ GUI features are implemented and tested
• 📖 Documentation included on the official doc webpage
• Run on local conda, virtual env and docker

v0.0.1

Initial release.
Set up initial GUI, platforms, env and three basic RL algorithms