The classic game of Pacman built with Pygame, provided also with a Reinforcement Learning environment.

• Quick Start
  • Game
  • RL Enviroment

Install the requirements

pip install -r requirements.txt

Run the Game with the classic maze

python main.py -lay classic -snd

Run the Game without music or sounds

python main.py -lay classic

Run the game with others option

usage: main.py [-h] [-lay LAYOUT] [-snd] [-stt]

Argument for the Pacman Game

optional arguments:
  -h, --help            show this help message and exit
  -lay LAYOUT, --layout LAYOUT
                        Name of layout to load in the game
  -snd, --sound         Activate sounds in the game
  -stt, --state         Display the state matrix of the game

The PacmanEnv class extends the gym.Env class, so if you already know how to use the open ai gym, the api is the same.

Here's a little example:

import gym

env = gym.make('pacman-v0', layout=self.layout, frame_to_skip=10)

for episode in range(episodes):
    env.reset()
    for i in range(max_steps):
        action = env.action_space.sample()
        obs, rewards, done, info = env.step(action)
        
        if done: 
            break

The src.env folder provides also an abstract class that you can use to make your own AI agent. You can use it to make your own agent, train it and directly plug into the game and see how will perform.

Here's how you can use it:

from src.env.agent import Agent

class MyAgent(Agent):
    name = 'my_agent'

    def __init__(self):
        pass

    def act(self, state, **kwargs):
    """
    The code that return the action to take
    """
        pass
    
    def train(self, **kwargs):
    """
    Your code to train the agent
    """
        pass

And after you're done with the training you can simply plug it into the game:

def run_agent(layout: str):
    agent = MyAgent(layout=layout)
    controller = Controller(layout_name=layout, act_sound=True, act_state=True, ai_agent=agent)
    controller.load_menu()

For more examples check out the examples folder.

• refactor everything using ECS
• implement fruit
• flashing power pellet
• state matrix in another screen
• Provide an RL Environment so an AI agent can be trained

MIT

Paolo D'Elia