KraTUZen/Vizdoom-Doom-Agent

🕹️ PPO Agent on ViZDoom Doom Agent

This repository contains a trained Proximal Policy Optimization (PPO) agent that plays the ViZDoom Doom environment, built using the ViZDoom framework (vizdoom.cs.put.edu.pl) (vizdoom.cs.put.edu.pl in Bing) (bing.com in Bing) and integrated with reinforcement learning libraries.

---

📊 Model Card

Model Name: ppo-ViZDoom-DoomAgent
Environment: ViZDoom Doom
Algorithm: PPO (Proximal Policy Optimization)
Performance Metric:

• Learns to shoot moving enemies in the Doom scenario
• Demonstrates convergence to a stable combat policy

---

🚀 Usage

from huggingface_hub import hf_hub_download
import gym

# Load the trained PPO model
model_path = hf_hub_download(
    repo_id="KraTUZen/Vizdoom-Doom-Agent",
    filename="model.pkl"
)

# Initialize environment
env = gym.make("Vizdoom-v0")

---

🧠 Notes

• The agent is trained using PPO, an on‑policy algorithm effective in visual, high‑dimensional tasks.
• The environment is ViZDoom Doom, where the agent must shoot enemies while navigating an arena.
• The serialized policy is stored in model.pkl.

---

📂 Repository Structure

• model.pkl → Trained policy weights
• README.md → Documentation and usage guide
• configs/ → Training configuration files
• scripts/ → Training and evaluation scripts

---

✅ Results

• The agent learns to aim and shoot effectively.
• Shows stable convergence using PPO, balancing exploration and exploitation.

---

🔎 Environment Overview

• Observation Space: Visual pixel‑based input (first‑person view)
• Action Space: Discrete (move, turn, shoot)
• Objective: Eliminate enemies efficiently
• Reward: Positive reward for hitting enemies, penalties for missed shots or damage

---

📚 Learning Highlights

• Algorithm: PPO (Proximal Policy Optimization)
• Update Rule: Clipped surrogate objective ensures stable updates
• Strengths: Handles pixel‑based visual input and sparse rewards
• Limitations: Training requires significant compute due to visual complexity