Capture The Flag Gridworld

This repository contains the implementation of multi-agent capture the flag in 2D, gridworld. Based on pycolab

Basic APIs

The api for the environment is quite simple, since it is similar to OpenAI's gym. For now we assume fully-observability (yes, the CTF will be in partial observability)

After creating the environment

env = FullObservableMultiAgentEnvironment(
    GAME_ART,
    {'P': PlayerSprite1, 'Q':PlayerSprite2},
    COLOR_MAP
)

We can run the environment by inputing a list

obs, reward, done = env.reset()
obs, reward, done = env.step([3, 1])

where each entries corresponds to action of each agent in this case player number 0 playing action 3, and player number 2 playing action 1, which will get out reward in a list of tuple, which indicates which agents got, which reward.

[(1, 0.0), (0, 100.0)]

---

For the Partial Observable Multi-Agent Environment (the window size is given) and all the same API

env = PartialObservableMultiAgentEnvironment(
    GAME_ART,
    {'P': PlayerSprite1, 'Q':PlayerSprite2},
    COLOR_MAP,
    window_size=3
)

Adding/Specify Player Behavior

In order to create the game, we will have to define the behavior of the players first (by using our normal controller -- Player)

class PlayerSprite2(Player):
    def __init__(self, corner, position, character):
        self.playerNum = 1
        super().__init__(corner, position, character)

    def behavior(self, actions, board, the_plot):
        if ord('$') == board[self.position]:
            self.assignReward(the_plot, 100.0)
        else:
            self.assignReward(the_plot, 0.0)

In this case, we have to always assign the player sprite a number (MUST starting from 0) so that we can order agent for observation output. All agents that inherited the Player class will have to following action

• Action 0 -- Up
• Action 1 -- Right
• Action 2 -- Down
• Action 3 -- Left

behavior method is used to define how the reward is assigned and when the game will ends (might also has a control over the dynamic of the game)

The agent, in the above example, gets reward 100 (tuple arrange in the term of player_num(tag) and reward) if it touches '$' on the map, and every step reward is zero

Tests

To run tests on picking up flag and returning to base, and receiving reward run Test two players grabbing flag at the same time

pytest testes

Hacking Here:

# Hack to allow absolute import from the root folder
if __name__ == "__main__" and __package__ is None:
    from sys import path
    from os.path import dirname as dir

    path.append(dir(path[0]))
    __package__ = "gridworld"