Atari

Implementation of Atari reinforcement learning algorithm

Technical Details

• Framework: PyTorch
• Environment: Atari
• Category: Other

This directory contains Proximal Policy Optimization (PPO) implementations for training agents on classic Atari games using PyTorch and Gymnasium.

🎮 Overview

This implementation features vectorized PPO training on Atari environments with proper preprocessing, frame stacking, and environment wrappers. The code includes both custom implementations and Stable Baselines3 benchmarks for comparison.

🏓 Trained Agents in Action

Pong

PPO agent playing Pong after 10M training steps - achieving consistent wins against the built-in AI opponent

Bowling

PPO agent playing Bowling after 10M training steps - demonstrating learned bowling strategies and consistent scoring

Boxing

PPO agent playing Boxing after 10M training steps - showing aggressive fighting strategies and effective combat techniques

🎯 Supported Games

• Pong (PongNoFrameskip-v4) - Classic paddle game
• Boxing (BoxingNoFrameskip-v4) - Fighting game
• Bowling (BowlingNoFrameskip-v4) - Bowling simulation

📁 Files

• atari-pong.py - PPO implementation for Pong
• boxing-atari.py - PPO implementation for Boxing
• atari-bowling.py - PPO implementation for Bowling
• sb3-atari-benchmark.py - Stable Baselines3 benchmark comparison
• images/ - Contains training videos and results

🚀 Features

Core PPO Implementation

• Vectorized Training: 8 parallel environments for efficient data collection
• Generalized Advantage Estimation (GAE): λ = 0.95 for bias-variance tradeoff
• Clipped Surrogate Objective: Prevents destructive policy updates
• Value Function Clipping: Stabilizes critic training
• Gradient Clipping: Prevents exploding gradients (max_grad_norm = 0.5)

Atari-Specific Features

• Frame Preprocessing: Grayscale conversion and resizing to 64x64
• Frame Stacking: 4 consecutive frames as input
• Standard Atari Wrappers:
  • NoopResetEnv: Random no-op actions at episode start
  • MaxAndSkipEnv: Frame skipping and max pooling
  • EpisodicLifeEnv: Treats life loss as episode end
  • FireResetEnv: Automatically fires at episode start
  • ClipRewardEnv: Clips rewards to [-1, 1]

Network Architecture

CNN Feature Extractor:
- Conv2d(4, 32, kernel=8, stride=4) + ReLU
- Conv2d(32, 64, kernel=4, stride=2) + ReLU  
- Conv2d(64, 64, kernel=3, stride=1) + ReLU
- Flatten
- Linear(64*7*7, 512) + ReLU

Actor Head: Linear(512, action_space)
Critic Head: Linear(512, 1)

⚙️ Hyperparameters

🏃‍♂️ Quick Start

Prerequisites

pip install torch gymnasium ale-py stable-baselines3 wandb opencv-python imageio

Training

# Train on Pong
python atari-pong.py

# Train on Boxing
python boxing-atari.py

# Train on Bowling
python atari-bowling.py

# Run SB3 benchmark
python sb3-atari-benchmark.py

Configuration

Edit the Config class in each file to modify hyperparameters:

• total_timesteps: Total training steps (default: 10M)
• env_id: Environment name
• lr: Learning rate
• num_envs: Number of parallel environments
• use_wandb: Enable Weights & Biases logging

📊 Results

Training Videos

Training videos are saved in the images/ directory:

• pong.mp4 - Trained Pong agent gameplay
• boxing.mp4 - Trained Boxing agent gameplay
• bowling.mp4 - Trained Bowling agent gameplay

Detailed Training Reports

📈 PPO Atari Pong - Training Report

🎳 PPO Atari Bowling - Training Report

🥊 PPO Atari Boxing - Training Report

The W&B reports include:

• Training curves and learning progression
• Hyperparameter sweeps and optimization
• Performance metrics and comparisons
• Loss functions and gradient analysis
• Real-time training monitoring

📈 Performance

The implementation typically achieves:

• Pong: 15-20 average reward after 10M steps
• Boxing: 80-95 average reward after 10M steps
• Bowling: 40-60 average reward after 10M steps

🔧 Technical Details

Environment Preprocessing

1. Observation: 210x160x3 RGB frames → 64x64 grayscale
2. Frame Stacking: 4 consecutive frames
3. Reward Clipping: Rewards clipped to [-1, 1]
4. Life Management: Episode ends on life loss

Training Loop

1. Rollout Collection: Collect trajectories from vectorized environments
2. Advantage Estimation: Compute GAE advantages
3. Policy Update: Multiple PPO epochs with minibatch updates
4. Value Update: Train critic with clipped value loss

🎯 Key Insights

• Vectorization: Dramatically improves sample efficiency
• Frame Stacking: Provides temporal information for decision making
• Proper Preprocessing: Essential for stable Atari training
• Clipping: Prevents destructive policy updates
• GAE: Reduces variance in advantage estimation

🔍 Monitoring

• Tensorboard: Real-time training metrics
• Weights & Biases: Experiment tracking and visualization
• Video Recording: Periodic agent gameplay videos
• Console Logging: Episode rewards and training progress

📚 References

• Proximal Policy Optimization
• Human-level control through deep reinforcement learning
• Generalized Advantage Estimation

💡 Tips for Success

1. Sufficient Training Time: Atari games require 10M+ steps
2. Proper Preprocessing: Use standard Atari wrappers
3. Stable Learning Rate: 2.5e-4 works well for most games
4. Monitor Training: Watch for policy collapse or instability
5. Vectorization: Use multiple environments for efficiency

Source Code

📁 GitHub Repository: Atari (PPO Atari)

View the complete implementation, training scripts, and documentation on GitHub.