Monopoly RL Advanced

A compact, educational Monopoly simulator with Deep Q-Learning (DQN), richer game rules, multi-item trading, scripted opponents, and detailed logging. This project is designed for reinforcement learning research and experimentation, not for full rules-accurate gameplay.

---

Features

• Fuller Monopoly Rules:
  • Chance & Community Chest events
  • Jail (go to jail, get out by rolling doubles or paying)
  • Utilities & Railroads
  • Auctions when a player declines to buy
  • Simplified mortgaging (auto-mortgage lowest-value properties if cash < 0)
• Trading:
  • Multi-item (k-for-m) trades with optional cash sweetener
  • $100 premium for trades that complete a set for the receiver
  • Stochastic "lean to accept" behavior near fair trades
  • Safety checks for valid trades
• Agent:
  • DQN (MLP) with target network
  • Epsilon-greedy exploration with decay
  • Experience replay buffer
• Opponents:
  • Toggle between scripted bots (aggressive buyer, conservative, builder) or self-play pool
• Evaluation & Logging:
  • Tracks landing distribution, win rates, average wealth
  • CSV logs and weights checkpoints
• CLI Flags:
  • Control episodes, evaluation frequency, opponent type, and save directory

---

Quick Start

Requirements

• Python 3.7+
• No external dependencies (uses only numpy, argparse, csv, etc.)

Usage

python monopoly_rl_advanced.py --episodes 5000 --eval-every 250 --opponents scripted --save-dir ./mono_adv

Arguments:

• --episodes: Number of training episodes (default: 1000)
• --eval-every: Evaluation interval (default: 100)
• --opponents: Opponent type (scripted or selfplay)
• --save-dir: Directory for logs and weights

---

How It Works

Game Mechanics

• Board: 40 tiles, including properties, railroads, utilities, taxes, jail, chance/chest, free parking, and go-to-jail.
• Players: 4 per game, each with cash, properties, houses, and mortgage status.
• Actions: Skip, Buy, Build, Offer Trade, Pay to Leave Jail.
• Trading: Agents can propose multi-property trades with cash, evaluated stochastically for fairness.
• Mortgaging: If cash drops below zero, properties are auto-mortgaged until solvency or bankruptcy.

RL Agent

• Observation: Encodes ownership, houses, mortgages, position, cash, jail status, and completed sets.
• DQN: Simple MLP (implemented in numpy), trained via experience replay and target network updates.
• Exploration: Epsilon-greedy, decaying over time.

Opponents

• Scripted Bots: Aggressive, conservative, and builder styles.
• Self-Play: All agents use the same DQN policy.

Logging & Evaluation

• CSV Log: Records episode, winner, moves, agent epsilon, average wealth, agent wealth, evaluation win rate, and average moves.
• Weights: Saved every evaluation interval and at the end.

---

Output Files

• training_log.csv: Episode-by-episode stats.
• weights_ep{N}.json: DQN weights at evaluation checkpoints.
• weights_final.json: Final trained weights.

---

Limitations

• Not a full Monopoly engine: Some rules are simplified for RL speed and clarity.
• No GUI: CLI only.
• No external RL libraries: All neural network code is pure numpy.

---

Extending

• Add more sophisticated agents or opponent strategies.
• Tune DQN architecture or hyperparameters.
• Integrate with external RL libraries for advanced experimentation.

---

License

This project is for educational and research use. See LICENSE for details.

---

Credits

Vibecoded

---

Contact

For questions or suggestions, open an issue or contact chatgpt@asu.edu (Yes this email is valid).