Palace Card Game Implementation

Watch the Training and Test Run Video

Detailed explanation of the code is available at: https://castle-ai-eight.vercel.app/

This repository contains two implementations of the Palace card game: a traditional player-vs-computer version (main.py) and an AI-powered version using Deep Q-Learning (palace_dqn.py).

Overview

Palace is a card game where players aim to get rid of all their cards by playing them onto a central pile. Each player starts with:

• 3 face-down cards (unknown)
• 3 face-up cards (visible to all)
• 3 cards in hand

Files

main.py

A traditional implementation of the Palace card game featuring:

• Human vs Human gameplay
• Human vs Computer gameplay (with basic AI)
• Complete rule implementation including:
  • Special card effects (2, 7, 10)
  • Hand management
  • Face-down card mechanics
  • Pile pickup rules

Key Features:

• Interactive command-line interface
• Card validation and rule enforcement
• Flexible game modes
• Deck management and card distribution
• Turn-based gameplay system

palace_dqn.py

An advanced implementation using Deep Q-Learning Neural Networks featuring:

• Sophisticated AI agents that learn through gameplay
• Deep Q-Network (DQN) architecture with:
  • Input Layer: 91 neurons
  • Hidden Layer 1: 128 neurons (ReLU)
  • Hidden Layer 2: 64 neurons (ReLU)
  • Output Layer: 3 neurons

Key Features:

• State representation (91-dimensional vector)
• Experience replay memory
• ε-greedy exploration strategy
• Reward system for reinforcement learning
• PyTorch implementation of neural networks

Requirements

numpy pytorch

Usage

Traditional Game (main.py)

Follow the prompts to:

1. Choose game mode (Human vs Human or Human vs Computer)
2. Play cards by entering suit and rank
3. Choose to pick up pile when necessary

AI Training (palace_dqn.py)

This will:

1. Initialize two AI agents
2. Train them through 1000 episodes
3. Display final test game results

Game Rules

1. Players must play cards of equal or higher rank than the top card

2. Special cards:

   • 2: Resets pile, play again
   • 7: Next player must play 7 or lower
   • 10: Burns pile, play again

3. Card hierarchy:

   • 2-10: Face value
   • Jack: 11
   • Queen: 12
   • King: 13
   • Ace: 14
   • Joker: 15

AI Implementation Details

State Representation

• 15 dimensions × 3 types for Player 1
• 15 dimensions × 3 types for Player 2
• 15 dimensions for pile top card
• 1 dimension for special rules

Reinforcement Learning Parameters

• Learning Rate (α): 0.001
• Discount Factor (γ): 0.99
• Initial Exploration Rate (ε): 1.0
• Exploration Decay: 0.995
• Minimum Exploration: 0.01

Reward Structure

• +10: Winning the game
• +1: Valid play
• -5: Invalid play
• -10: Picking up pile