kymnasium 1.0.7.dev2

The collection of reinforcement learning environments developed for the Artificial Intelligence course at the Department of Computer Science and Engineering, Kangwon National University.

KNU Gymnasium, or Kymnasium for Reinforcement Learning Environments

Welcome to Kymnasium! This project is a collection of reinforcement learning environments developed for the Artificial Intelligence course at the Department of Computer Science and Engineering, Kangwon National University.

About the Project

Kymnasium provides a simple and effective platform for students to learn, implement, and test reinforcement learning algorithms. The environments are designed to be straightforward, allowing you to focus on the core concepts of RL. This project is inspired by Gymnasium.

Environments

• Alkkagi: A Korean traditional game where the objective is to flick your stones to knock the opponent's stones off the board.
• Avoid Blurp: An environment where the agent must learn to avoid free-fall enemies.
• Grid Adventure: A classic grid-world environment where the agent navigates a maze to reach a goal.
• Grid Survivor: A grid-world environment where the agent must survive for as long as possible.

Getting Started

Installation

pip install kymnasium

Implement Your Agent

To train your own agent, you need to override 'kymnasium.Agent' and implement three methods as below:

from src import kymnasium as kym


# Your agent logic goes here
class YourAgent(kym.Agent):
    def act(self, observation: any, info: dict):
        # Replace this with your agent's action selection logic return env.action_space.sample()
        pass

    @classmethod
    def load(cls, path: str) -> 'kym.Agent':
        # Load a pre-trained agent
        pass

    def save(self, path: str):
        # Save the trained agent
        pass

Basic Training loop

import gymnasium as gym


# Train the agent for 100 episodes
EPISODES = 100

# Path for saving your agent
PATH_AGENT = './agent.pkl'
agent = YourAgent()

# Create the environment
env = gym.make(
    id="kymnasium/GridAdventure-FullMaze-26x26-v0", # Environment ID
    render_mode='rgb_array', # or 'human',
    obs_type='custom', # or 'image'
    bgm=False # or True for playing background music
)
for _ in range(EPISODES):
    observation, info = env.reset()
    done = False
    while not done: 
        action = agent.act(observation, info) 
        observation, reward, terminated, truncated, info = env.step(action) 
        if terminated or truncated: 
            done = True
        # Here writes any training logic

# Close the environment        
env.close()

# Save your agent
agent.save(PATH_AGENT)

Live Evaluation of Your Agent

from src import kymnasium as kym

evaluator = kym.LocalEvaluator(
    env_id="kymnasium/GridAdventure-FullMaze-26x26-v0",  # Environment ID
    agent=YourAgent.load(PATH_AGENT),  # Your trained agent
    render_mode='human',  # 'render_mode' should be 'human' for live evaluation
    obs_type='custom',  # or 'image'
    bgm=True  # 'bgm' should be 'True' for live evaluation
)

evaluator.run()

Manual Play

If you want to manually play the environment, see below:

from grid_adventure import ManualPlayWrapper

agent = ManualPlayWrapper(
    env_id='kymnasium/GridAdventure-FullMaze-26x26-v0',
    render_mode='human',  # 'render_mode' should be 'human' for manual play
)
agent.play()