jatinror/frozen-lake-q-agent

Frozen-Lake-v1 Q-Learning Agent

This repository contains a trained Q-learning agent for the Frozen-Lake-v1 environment from Gymnasium.

The agent is trained to navigate the frozen lake, avoid holes, and reach the goal efficiently. It uses a Q-table stored in model.zip.

---

Environment

• Name: FrozenLake-v1
• Observation Space: Discrete (depends on lake size, e.g., 6x6 → 36 states)
• Action Space: Discrete 4 (LEFT, DOWN, RIGHT, UP)
• Map: Randomized frozen lake with holes (H) and goal (G)
• Slippery: True (agent may slide on ice)

---

Algorithm

• Type: Q-learning (tabular)
• Hyperparameters:
  • Learning rate: 0.7
  • Discount factor (gamma): 0.95
  • Exploration (epsilon): decaying from 1.0 → 0.01
• Episodes: 8000+
• Max Steps per Episode: 200

---

Model

The trained Q-table is saved as:

Download model.zip

Inside model.zip, you will find q_table.npy.

---

How to Load the Model

import numpy as np

# Load Q-table
q_table = np.load("model.zip", allow_pickle=True)

# Example usage
state = 0  # starting state
action = np.argmax(q_table[state, :])