ML_Final_Project_Rock_Paper_Scissors_Game

Introduction

In the culminating project for our Machine Learning course, we were tasked with developing a machine learning agent to compete in the classic game of Rock-Paper-Scissors against a randomly playing opponent. The objective was to design an intelligent agent capable of recognizing the opponent's move and strategically countering it to secure a win. To achieve this, we employed a neural network model to empower our agent's decision-making capabilities, pitting it against an opponent that makes random choices

Agent Model

Our CNN model for the Rock-Paper-Scissors game is designed with the following key layers:

Convolutional Layers:

• First Layer: 32 filters (3x3).
• Second Layer: 64 filters (3x3).
• Activation: ReLU (Rectified Linear Unit) for introducing non-linearity.

Max Pooling Layers:

• Pool Size: 2x2, reducing spatial dimensions and computational load.

Flatten Layer:

Converts pooled feature maps into a single vector, serving as input to the dense layers.

Dense Layers:

• One dense layer with 128 neurons.
• Output layer with 3 neurons (for rock, paper, scissors) using softmax activation.

Model Advantages

• Feature Learning: Efficiently learns spatial features from images, crucial for distinguishing hand signs.
• Parameter Efficiency: Fewer parameters due to shared weights, enhancing performance.
• Robustness: Handles spatial variations effectively, ideal for recognizing hand signs irrespective of their positioning. This architecture is tailored for the task of recognizing and classifying hand signs in the Rock-Paper-Scissors game, leveraging CNN's strengths in image processing and pattern recognition.

Model Training and Validation

Our model is trained and validated on the Rock-Paper-Scissors dataset sourced from Kaggle: https://www.kaggle.com/datasets/drgfreeman/rockpaperscissors:

• Resizing: Images are reshaped to 32x32 to reduce model complexity.
• Color Transformation: RGB images are converted to grayscale.
• Normalization: Pixel values are normalized to the [0, 1] range by dividing each pixel by 255.

Game Simulation

Our agent competes against a "Random Agent" in a series of N rounds in the Rock-Paper-Scissors game.

• Betting Mechanism: Each round involves a bet of €1.
• Winning: If our agent wins, it earns €2.
• Tie: In a tie, the agent gets a refund of €1.
• Losing: If our agent loses, it forfeits €1.

Random Agent Strategy

• Image Selection: Randomly chooses from 2100 images (rock, paper, or scissors).
• Image Manipulation:
  • Applies a vertical flip with a 50% probability.
  • Applies a horizontal flip with a 50% probability.
  • Adds random noise to each pixel (mean 0, standard deviation 5% of the maximum pixel value).

Model's Performance on test set

Conclusion

Our AI Agent for the Rock-Paper-Scissors game showcases the capabilities of machine learning in both strategic gameplay and image recognition. It excels in handling our current dataset; however, enhancing its adaptability to a wider range of real-world situations is an objective we are actively pursuing. Keep an eye out for future developments and enhancements