DinoAI

Overview

DinoAI is a reinforcement learning project where an AI agent learns to play the Chrome Dino game using Deep Q-Networks (DQN). Instead of relying on hardcoded logic, the agent learns optimal actions (jump, duck, or idle) through continuous interaction with the game environment.

The project demonstrates how neural networks can be used for decision-making in dynamic environments by maximizing rewards over time.

Key Features

Core Functionalities

• DQN-based Learning: Uses Deep Q-Learning for decision making
• Game Frame Processing: Captures and processes screen frames
• Action Prediction: Decides between jump, duck, or no action
• Reward System: Encourages longer survival and better timing
• Two Versions:
  • Basic Model (~1 FPS): Initial implementation
  • Optimized Model (3–4 FPS): Improved accuracy and smoother decisions
• Adaptive Behavior: Learns obstacle patterns over time

How It Works

1. Frame Capture → Game screen captured in real-time
2. State Processing → Frames converted into model input
3. Action Selection → Model predicts best action
4. Environment Interaction → Executes action in game
5. Reward Feedback → Positive reward for survival
6. Model Update → Q-values updated using DQN

Core Concept

At the core, the agent learns an optimal policy using Q-learning.

This helps the model estimate the best action based on expected future rewards.

Use Cases

• Reinforcement learning experimentation
• Game AI development
• Understanding agent-based learning systems
• Educational ML projects

Performance Comparison

• Basic Model --> ~1 FPS --> Slower, less accurate
• Optimized Model --> 3–4 FPS --> Better precision, smoother

Future Improvements

• Increase FPS using better optimization
• Use CNNs for improved visual feature extraction
• Hyperparameter tuning (epsilon decay, learning rate)
• Experience replay optimization
• Deploy as interactive demo