Skip to content

PROJECT_SUMMARY.md

Latest commit

 

History

History
93 lines (76 loc) · 3.65 KB

PROJECT_SUMMARY.md

File metadata and controls

93 lines (76 loc) · 3.65 KB

Connect 4 Computer Vision AI Project

Project Overview

This project investigates how computer vision distortions (blur, misclassification) affect AI decision-making in Connect 4 gameplay. We simulate CV errors and measure their impact on AI performance.

Key Components

Core Game Logic

  • game.py: Connect 4 game mechanics, board operations, win detection
  • ai.py: Minimax AI with alpha-beta pruning (depth 5-6)

AI Players

  • deterministic_ai.py: Basic rule-based AI (simple strategies)
  • enhanced_deterministic_ai.py: Advanced deterministic AIs with position scoring
    • WeakDeterministicAI: Simple center-out strategy
    • ModerateDeterministicAI: Basic rule following
    • StrongDeterministicAI: Uses scoring algorithm, deterministic choices

Experimental Frameworks

  • final_cv_experiment.py: Main experiment runner

    • Tests 5 accuracy levels: 100%, 98%, 96%, 94%, 90%
    • 30 games per accuracy level (150 total games)
    • Measures: win rates, illegal moves, game length, move optimality
    • Exports data to CSV for analysis

  • digital_experiment_framework.py: Advanced experimental framework

    • Ground truth board generation (early/mid/late game scenarios)
    • Controlled distortion application (blur + misclassification)
    • Comprehensive metrics collection

Experimental Design

CV Distortion Simulation

  1. Blur: Randomly removes pieces from the board (simulates unclear vision)
  2. Misclassification: Swaps piece colors (simulates color detection errors)
  3. Combined: Both distortions applied simultaneously

Accuracy Levels Tested

  • 100%: Perfect vision (control group)
  • 98%: 1% blur + 1% misclassification
  • 96%: 2% blur + 2% misclassification
  • 94%: 3% blur + 3% misclassification
  • 90%: 5% blur + 5% misclassification

Key Metrics

  • Win Rate: Percentage of games won by distorted vs perfect AI
  • Illegal Moves: Moves that violate game rules due to distorted vision
  • Game Length: Number of moves per game
  • Move Optimality: Quality of decisions compared to oracle AI
  • Move Time: Decision-making speed

Key Findings

  • As CV accuracy decreases, AI performance degrades significantly
  • 100% accuracy: AI wins consistently with 0 illegal moves
  • Lower accuracies: Dramatic drop in win rate, increase in illegal moves
  • Game lengths become shorter as distorted AI makes poor decisions

Data Output

Results exported to CSV with columns:

  • Trial number, accuracy percentage, distortion levels
  • Board index, starting player, moves made
  • Win/loss outcomes, illegal move counts
  • Game length, move times, optimality scores

Usage

# Run main experiment
python final_cv_experiment.py

# Run advanced framework
python digital_experiment_framework.py

Technical Details

  • Python 3.11 with numpy for board operations
  • Minimax with alpha-beta pruning for strategic AI
  • Deterministic AIs ensure reproducible experiments
  • CSV export for Excel/statistical analysis

Files Structure

CS_EE_Project/
├── game.py                           # Core Connect 4 logic
├── ai.py                            # Minimax AI implementation  
├── deterministic_ai.py              # Basic deterministic AI
├── enhanced_deterministic_ai.py     # Advanced deterministic AIs
├── final_cv_experiment.py           # Main experiment runner
├── digital_experiment_framework.py  # Advanced experiment framework
├── *.csv                           # Experimental results
└── __pycache__/                    # Python cache files

This project demonstrates how computer vision accuracy directly impacts AI decision-making quality in strategic games, with applications to robotics and automated systems.