tankwar-env 1.0.1

MultiAgent gym environment for reinforcement learning

Installation

pip install tankwar-env

or

git clone https://github.com/khoda81/tankwar.git
cd tankwar
pip install -e

or

pip install git+https://github.com/khoda81/tankwar.git

Keyboard

• Environment:
  • Esc : set env.done to True (end episode on next step)
  • F : toggle limited frame rate
  
  
• HumanAgent:
  • W A S D : movement
  • Q E : rotate turret
  • Space : toggle shooting
  • Left Mouse Button : start shooting when pressed and stop shooting when released
  
  
• Script:
  • G : toggle window update (disabling window update will blur window and increase performance)

Basic Usage

import pygame

from tankwar.agents import HumanAgent, RandomAgent
from tankwar.envs import TankWarEnv


def main():
    random_agents = 2

    # create environment
    env = TankWarEnv(random_agents + 1, shape=(200, 200))

    # create agents
    agents = [HumanAgent(env)] + [RandomAgent(env) for _ in range(random_agents)]

    # initialize a window with the height of 200
    # width is calculated based on env.shape
    # limit frame rate to 60 if a human is playing
    env.init_window(600, True)
    done = False

    # reset environment
    observations = env.reset()
    display_to_human = True

    while True:
        if display_to_human:
            env.render("human")  # render to screen
        else:
            # if render is not called window events are not processed either
            # so, manually processing window events
            env.process_window_events()

        actions = [
            agent.act((None, None), 0, done)
            for agent in agents
        ]

        observations, rewards, done, info = env.step(actions, print_fps=True)

        # if g is pressed screen is toggled
        # disabling screen will make rendering faster
        # because copying frame to window takes time
        for event in env.events:
            if event.type == pygame.KEYDOWN and event.key == pygame.K_g:
                display_to_human = not display_to_human

                if not display_to_human:
                    env.blur_window()

        if done:
            break


if __name__ == "__main__":
    main()

Main Usage

import time

import pygame
import torch
from torch.nn.functional import interpolate
from tqdm import tqdm

from tankwar.agents import HumanAgent, RandomAgent
from tankwar.envs import TankWarEnv

MAX_EPISODE_STEPS = 5000


def main():
    random_agents = 2
    human_agent = False  # whether or not a human is currently playing
    display_to_human = human_agent

    # create environment
    env = TankWarEnv(random_agents + human_agent, shape=(200, 200))
    w, h = env.shape

    # create agents
    agents = (
            [HumanAgent(env)] * human_agent +
            [RandomAgent(env) for _ in range(random_agents)])

    # initialize a window with the height of 200
    # width is calculated based on env.shape
    # limit frame rate to 60 if a human is playing
    env.init_window(h, human_agent)
    padded_frame = torch.zeros(3, w + 200, h + 200)

    observations = env.reset()
    done = False

    rewards = [0] * env.n

    s = time.time()
    for _ in tqdm(range(MAX_EPISODE_STEPS), unit="step"):
        # render to screen
        if display_to_human:
            env.render("human")

        # convert frame to pytorch tensor with shape (3, height, width)
        # frames are cached and will be rendered once per step
        frame = env.render("rgb_array")
        frame_torch = (torch.from_numpy(frame) / 255).permute(2, 0, 1)

        # each frame will be down sampled to (w, h)
        if env.window_scale != 1:
            frame_torch = interpolate(frame_torch.reshape(1, *frame_torch.shape), (h, w))[0]
        # with 100 pixels padding on each side:
        padded_frame[:, 100:-100, 100:-100] = frame_torch

        # image that each agent can see
        images = []
        for i, tank in enumerate(env.tanks):
            x, y = tank.body.position.int_tuple
            images.append(padded_frame[:, y:y + 200, x:x + 200])

        actions = [
            agent.act((observation, image), reward, done)
            for agent, observation, image, reward
            in zip(agents, observations, images, rewards)
        ]

        observations, rewards, done, info = env.step(actions)

        # if g is pressed screen is toggled
        # disabling screen will make rendering faster
        # because copying frame to window takes time
        for event in env.events:
            if event.type == pygame.KEYDOWN and event.key == pygame.K_g:
                display_to_human = not display_to_human

                if not display_to_human:
                    env.blur_window()

        if done:
            break

    print(f"Episode took {time.time() - s:.2f} seconds.")


if __name__ == "__main__":
    main()