This repository contains two small, self-contained Reinforcement Learning (RL) projects built to demonstrate the fundamentals of Q-Learning, one of the most widely used algorithms for teaching agents to make decisions through experience.
Both examples are minimal yet illustrative — designed for learning, experimentation, and portfolio display.
File: simple_rl/simple_q_learning.py
A minimal environment with three states — start, candy, and fire — and two possible actions (left, right).
The goal is for the agent to learn that moving right from the start leads to the “candy” reward (+1) while moving left leads to “fire” (–1).
- Basic Q-table initialization and updates
- Reward shaping (penalizing long paths)
- Learning rate (α) and discount factor (γ) tuning
- Simple environment modeling
Each episode starts in the start state.
The agent randomly chooses actions, receives rewards, and updates its Q-table using the formula:
After several episodes, the agent learns to consistently choose the optimal action (right toward candy).
File: gridworld_rl/gridworld_q_learning.py
A classic GridWorld setup where the agent starts in the top-left corner (state 0) and learns to reach the bottom-right goal (state 15) on a 4×4 grid.
The environment uses small negative rewards (–0.01) for movement and a positive reward (+1) for reaching the goal.
- 2D state-space navigation
- Epsilon-greedy exploration strategy
- Reward-based policy optimization
- Learned policy visualization using arrows (
↑,→,↓,←)
The agent starts at state 0 and repeatedly explores random paths.
Over time, it learns which actions lead to the goal fastest by maximizing cumulative rewards.
The resulting Q-table is converted into a visual policy grid showing the best move for each cell.
Clone the repository:
git clone https://github.com/<your-username>/Q-Learning-Mini-Projects.git
cd Q-Learning-Mini-Projects
Run either project:
python3 simple_rl/simple_q_learning.py
python3 gridworld_rl/gridworld_q_learning.py
You’ll see printed output of:
The final Q-table
The optimal policy (best action per state)
The path the agent takes to reach its goal
Requirements:
Python 3.9+
Dependencies:
numpy (for GridWorld)
Built-in libraries: random
Install required packages:
pip install numpy