Descent Visualisers provides a suite of interactive labs designed to enhance understanding of various Machine Learning algorithms through engaging visualizations. This project focuses on two primary areas of study: Gradient Descent and Reinforcement Learning.

Gradient Descent Visualiser

The Gradient Descent Visualiser is a dynamic tool that allows users to observe how different gradient descent algorithms operate on complex loss surfaces in real time. Key features include:

• Exploration of optimizers such as Vanilla SGD, Momentum, Adagrad, RMSProp, and Adam.
• Real-time 3D visualizations facilitated by Plotly.
• The ability to input custom loss surfaces by utilizing various mathematical functions.
• Adjustable hyperparameters including learning rate, momentum, decay, and beta values.
• Functionalities to play, pause, reset animations, and reinitialize points.

Preview Demonstrations

3D Loss Surface	Different Optimizers	Custom Surface

Reinforcement Learning Visualiser

The Reinforcement Learning Visualiser explores the behavior of various Reinforcement Learning algorithms such as Q-learning (QL), Deep Q-learning (DQL), and Proximal Policy Optimization (PPO) within custom simulated environments.

Environments & Features

Maze Environment

• Utilizes tabular Q-Learning for training.
• Adjustable hyperparameters to customize learning dynamics.
• Editable configurations for unique environment setups.
• Visualized training steps to observe learning progression.
• Play mode for interacting with trained agents.

CartPole Environment

• Trained using both DQL and PPO strategies.
• Adjustable hyperparameters for optimal training experiences.
• Configuration of environment parameters for tailored scenarios.
• Visual aids to display training steps and agent behavior.
• Play mode to observe the performance of trained agents.

Preview Demonstrations

Maze Environment	CartPole Environment

Live Demo

Access the live demo at: Descent Visualisers Live Demo

Engaging with these visualizations enables a deeper comprehension of algorithm behavior and optimizes the learning process in machine learning and reinforcement learning domains.