---
library_name: numpy
tags:
- Taxi-v3
- reinforcement-learning
- q-learning
- custom-implementation
model-index:
- name: Q-Learning
  results:
  - task:
      type: reinforcement-learning
      name: reinforcement-learning
    dataset:
      name: Taxi-v3
      type: Taxi-v3
    metrics:
    - type: mean_reward
      name: mean_reward
      value: 7.92 +/- 2.60
      verified: false
---

# 🚖 Q-Learning Agent for Taxi-v3

This is a trained **Q-Learning agent** for the **Taxi-v3** environment using a **tabular approach**.

## Developer
**Vishand S (@Vishand03)**

## Frameworks
- Python  
- NumPy  
- Gymnasium

## Training Details
- Algorithm: Q-Learning  
- Episodes: 2,000,000  
- Max Steps per Episode: 200  
- Learning rate (α): 0.1  
- Discount factor (γ): 0.99  
- Exploration: Epsilon-greedy  
- Epsilon decay: 0.0005  
- Mean Reward: ~7.92 ± 2.60  

---

## 🛠 Usage

```python
import gymnasium as gym
import pickle
from huggingface_hub import hf_hub_download

# -------------------------
# Load pretrained model
# -------------------------
model_file = hf_hub_download("Vishand03/q-Taxi-v3", "q-learning.pkl")
with open(model_file, "rb") as f:
    model = pickle.load(f)

env = gym.make(model["env_id"])

# -------------------------
# Evaluate agent
# -------------------------
def greedy_policy(Qtable, state):
    return max(range(len(Qtable[state])), key=lambda a: Qtable[state][a])

total_rewards = []
for _ in range(model["n_eval_episodes"]):
    state, _ = env.reset()
    done = False
    episode_reward = 0
    while not done:
        action = greedy_policy(model["qtable"], state)
        state, reward, terminated, truncated, _ = env.step(action)
        episode_reward += reward
        done = terminated or truncated
    total_rewards.append(episode_reward)

mean_reward = sum(total_rewards) / len(total_rewards)
print(f"Mean Reward: {mean_reward:.2f}")