My Reinforcement Learning Journey
Interactive demonstrations of my progress in learning and implementing RL algorithms
Introduction
Follow my journey in learning reinforcement learning through interactive demonstrations. Starting with a simple gridworld environment, I'll be adding more complex environments and algorithms as I progress in my learning journey.
Learning Resources
Check out my interactive visualizations to understand key RL concepts:
- Bellman Equation Visualizer - An interactive tool to understand the fundamental equation of RL
- Interactive grid world with adjustable parameters
- Step-by-step visualization of value iteration
- Q-learning implementation with real-time updates
Gridworld Navigation
Deep Q-Network in Gridworld
An environment where an agent learns to navigate to randomly placed goals while avoiding obstacles. This demonstrates key concepts in deep reinforcement learning:
- Deep Q-Network (DQN) architecture
- Experience replay buffer
- Target network for stable learning
- Epsilon-greedy exploration
Implementation Details
The agent uses DQN with the following parameters:
- Learning rate: 0.001
- Discount factor (γ): 0.99
- Initial ε: 1.0, decay: 0.997, minimum: 0.01
- Reward structure: -0.01 per step, +1.0 for goal
- Neural network: 2-layer fully connected (64 units each)
- State: Normalized (x,y) coordinates
- Batch size: 64
Future Environments
As I continue my reinforcement learning journey, I'll be adding more environments and algorithms:
- CartPole: A classic control problem to learn deep Q-networks (DQN)
- Mountain Car: Exploring the challenges of sparse rewards
- Lunar Lander: Implementing more advanced algorithms like PPO
- Custom Environments: Creating my own environments for specific learning objectives
Check back regularly to see new demonstrations and implementations as I progress in my learning journey!