generals 0.5.1

Generals.io environment compliant with PettingZoo API standard powered by Numpy.

Generals.io RL

Installation • Getting Started • Customization • Environment

Generals.io is a real-time strategy game where players compete to conquer their opponents' generals on a 2D grid. While the goal is simple — capture the enemy general — the gameplay involves a lot of depth. Players need to employ strategic planning, deception, and manage both micro and macro mechanics throughout the game. The combination of these elements makes the game highly engaging and complex.

This repository aims to make bot development more accessible, especially for Machine Learning based agents.

Highlights:

• 🚀 Fast & Lightweight simulator powered by numpy (thousands of steps per second)
• 🤝 Compatibility with Reinforcement-Learning API standards 🤸Gymnasium and 🦁PettingZoo
• 🔧 Easy customization of environments
• 🔬 Analysis tools such as replays

📦 Installation

Stable release version is available through pip:

pip install generals

Alternatively, you can install latest version via git

git clone https://github.com/strakam/Generals-RL
cd Generals-RL
pip install -e .

Usage example (🤸 Gymnasium)

from generals.env import gym_generals
from generals.agents import RandomAgent, ExpanderAgent

# Initialize agents
agent = RandomAgent()
npc = ExpanderAgent()

# Create environment -- render modes: {None, "human"}
env = gym_generals(agent=agent, npc=npc, render_mode="human")
observation, info = env.reset()

done = False

while not done:
    action = agent.play(observation)
    observation, reward, terminated, truncated, info = env.step(action)
    done = terminated or truncated
    env.render(fps=6)

You can also check an example for 🦁PettingZoo or more extensive example here.

🚀 Getting Started

Creating your first agent is very simple.

• Start by subclassing an Agent class just like RandomAgent or ExpanderAgent.
• Every agent must have a name as it is his ID by which he is called for actions.
• Every agent must implement play(observation) function that takes in observation and returns an action.
• You can start by copying the Usage Example and replacing agent with your implementation.
• When creating an environment, you can choose out of two render_modes:
  • None that omits rendering and is suitable for training,
  • "human" where you can see the game roll out.

[!TIP] Check out Makefile and run some examples to get a feel for the game 🤗.

🎨 Custom maps

Maps are handled via Mapper class. You can instantiate the class with desired map properties, and it will generate maps with these properties for each run.

from generals.env import pz_generals
from generals.map import Mapper

mapper = Mapper(
    grid_dims=(10, 10),                    # Dimensions of the grid (height, width)
    mountain_density=0.2,                  # Probability of a mountain in a cell
    city_density=0.05,                     # Probability of a city in a cell
    general_positions=[(0,3),(5,7)],       # Positions of generals (i, j)
)

# Create environment
env = pz_generals(mapper=mapper, ...)

You can also specify map manually, as a string via options dict:

from generals.env import pz_generals
from generals.map import Mapper

mapper = Mapper()
env = pz_generals(mapper=mapper, ...)

map = """
.3.#
#..A
#..#
.#.B
"""

options = {'map' : map}

# Pass the new map to the environment (for the next game)
env.reset(options=options)

Maps are encoded using these symbols:

• . for cells where you can move your army
• # for mountains (terrain that can not be passed)
• A,B are positions of generals
• digits 0-9 represent cities with the cost calculated as (40 + digit)

🔬 Replays

We can store replays and then analyze them. Replay class handles replay related functionality.

Storing a replay

from generals.env import pz_generals

options = {"replay": "my_replay"}
env = pz_generals(...)
env.reset(options=options) # The next game will be encoded in my_replay.pkl

Loading a replay

from generals.replay import Replay

# Initialize Replay instance
replay = Replay.load("my_replay")
replay.play()

🕹️ Replay controls

• q — quit/close the replay
• r — restart replay from the beginning
• ←/→ — increase/decrease the replay speed
• h/l — move backward/forward by one frame in the replay
• spacebar — toggle play/pause
• mouse click on the player's row — toggle the FoV (Field of View) of the given player

🌍 Environment

🔭 Observation

An observation for one agent is a dictionary {"observation": observation, "action_mask": action_mask}.

The observation is a Dict. Values are either numpy matrices with shape (N,M), or simple int constants:

Key	Shape	Description
army	(N,M)	Number of units in a cell regardless of owner
general	(N,M)	Mask of cells that are visible to the agent
city	(N,M)	Mask saying whether a city is in a cell
visibile_cells	(N,M)	Mask indicating cells that are visible to the agent
owned_cells	(N,M)	Mask indicating cells controlled by the agent
opponent_cells	(N,M)	Mask indicating cells owned by the opponent
neutral_cells	(N,M)	Mask indicating cells that are not owned by agents
structure	(N,M)	Mask indicating whether cells contain cities or mountains, even out of FoV
owned_land_count	—	Number of cells an agent owns
owned_army_count	—	Total number of units of an agent over all cells
opponent_land_count	—	Number of cells owned by the opponent
opponent_army_count	—	Int representing total number of units owned by the opponent
is_winner	—	Whether agent won
timestep	—	Timestep

action_mask is a mask with shape (N,M,4) where value [i,j,d] says whether you can move from cell [i,j] in a direction d.

⚡ Action

Action is a tuple(pass, cell, direction, split), where:

• pass indicates whether you want to 1 (pass) or 0 (play).
• cell is an np.array([i,j]) where i,j are indices of the cell you want to move from
• direction indicates whether you want to move 0 (up), 1 (down), 2 (left), or 3 (right)
• split indicates whether you want to 1 (split) units (send half of them) or 0 (no split), which sends all possible units to the next cell.

[!TIP] You can see how actions and observations look like by printing a sample form the environment:

print(env.observation_space.sample())
print(env.action_space.sample())

🎁 Reward

It is possible to implement custom reward function. The default is 1 for winner and -1 for loser, otherwise 0.

def custom_reward_fn(observation, action, done, info):
    # Give agent a reward based on the number of cells they own
    return observation["observation"]["owned_land_count"]

env = generals_v0(reward_fn=custom_reward_fn)
observations, info = env.reset()