Skip to content

Latest commit

 

History

History
249 lines (186 loc) · 13.1 KB

README.md

File metadata and controls

249 lines (186 loc) · 13.1 KB

SMACv2 Documentation

Introduction

SMACv2 is an update to Whirl’s Starcraft Multi-Agent Challenge, which is a benchmark for research in the field of cooperative multi-agent reinforcement learning. SMAC and SMACv2 both focus on decentralised micromanagement scenarios in StarCraft II, rather than the full game. It makes use of Blizzard’s StarCraft II Machine Learning API as well as Deepmind’s PySC2. We hope that you will enjoy using SMACv2! More details about SMAC can be found in the SMAC README as well as the SMAC paper. SMAC retains exactly the same API as SMAC so you should not need to change your algorithm code other than adjusting to the new observation and state size.

If you encounter difficulties using SMACv2, or have suggestions please raise an issue, or better yet, open a pull request!

The aim of this README is to answer some basic technical questions and to get people started with SMACv2. For a more scientific account of the work of developing the benchmark, please read SMACv2 paper. Videos of learned policies are available on our website.

Differences To SMAC

SMACv2 makes three major changes to SMACv2: randomising start positions, randomising unit types, and changing the unit sight and attack ranges. These first two changes were motivated by the discovery that many maps in SMAC lack enough randomness to challenge contemporary MARL algorithms. The final change increases diversity among the different agents and brings the sight range in line with the true values in StarCraft. For more details on the motivation behind these changes, please check the accompanying paper, where these are discussed in much more detail!

Capability Config

All the procedurally generated content in SMACv2 is managed through the Capability Config. This describes what units are generated and in what positions. The presence of keys in this config tells SMACv2 that a certain environment component is generated or not. As an example, consider the below config:

capability_config:
    n_units: 5
    team_gen:
      dist_type: "weighted_teams"
      unit_types: 
        - "marine"
        - "marauder"
        - "medivac"
      weights:
        - 0.45
        - 0.45
        - 0.1
      exception_unit_types:
        - "medivac"
      observe: True

    start_positions:
      dist_type: "surrounded_and_reflect"
      p: 0.5
      n_enemies: 5
      map_x: 32
      map_y: 32

This config is the default config for the SMACv2 Terran scenarios. The start_positions key tells SMACv2 to randomly generate start positions. Similarly the team_gen key tells SMACv2 to randomly generate teams. The dist_type tells SMACv2 how to generate some content. For example, team generation has the key weighted_teams , where each unit type is spawned with a certain weight. In this case a Stalker is spawned with probability 0.45 for example. Don’t worry too much about the other options for now — they are distribution-specific.

All the distributions are implemented in the distributions.py file. We encourage users to contribute their own keys and distributions for procedurally generated content!

Random Start Positions

Random start positions come in two different types. First, there is the surround type, where the allied units are spawned in the middle of the map, and surrounded by enemy units. An example is shown below.

This challenges the allied units to overcome the enemies approach from multiple angles at once. Secondly, there are the reflect scenarios. These randomly select positions for the allied units, and then reflect their positions in the midpoint of the map to get the enemy spawn positions. For example see the image below.

The probability of one type of scenario or the other is controlled with the p setting in the capability config. The cones are not visible in the above screenshot because they have not spawned in yet.

Random Unit Types

Battles in SMACv2 do not always feature units of the same type each time, as they did in SMAC. Instead, units are spawned randomly according to certain pre-fixed probabilities. Units in StarCraft II are split up into different races. Units from different races cannot be on the same team. For each of the three races (Protoss, Terran, and Zerg), SMACv2 uses three unit types.

Race Unit Generation Probability
Terran Marine 0.45
Marauder 0.45
Medivac 0.1
Protoss Stalker 0.45
Zealot 0.45
Colossus 0.1
Zerg Zergling 0.45
Hydralisk 0.45
Baneling 0.1

Each race has a unit that is generated less often than the others. These are for different reasons. Medivacs are healing-only units and so an abundance of them leads to strange, very long scenarios. Colossi are very powerful units and over-generating them leads to battles being solely determined by colossus use. Banelings are units that explode. If they are too prevalent, the algorithm learns to hide in the corner and hope the enemies all explode!

These weights are all controllable via the capability_config . However, if you do decide to change them we recommend that you do some tests to check that the scenarios you have made are sensible! Weights changes can sometimes have unexpected consequences.

Getting Started

This section will take you through the basic set-up of SMACv2. The set-up process has changed very little from the process for SMAC, so if you are familiar with that, follow the steps as you usually would. Make sure you have the 32x32_flat.SC2Map map file in your SMAC_Maps folder. You can download the SMAC_Maps folder here.

First, you will need to install StarCraft II. On windows or mac, follow the instructions on the StarCraft website. For linux, you can use the bash script here. Then copy

Then simply install SMAC as a package:

pip install git+https://github.com/oxwhirl/smacv2.git

[NOTE]: If you want to extend SMACv2, you must install it like this:

git clone https://github.com/oxwhirl/smacv2.git
cd smacv2
pip install -e ".[dev]"
pre-commit install

If you tried these instructions and couldn’t get SMACv2 to work, please let us know by raising an issue.

We also added configs for the protoss, terran and zerg configs to the examples folder. Note that you will have to change the n_units and n_enemies config to access the different scenarios. For clarity, the correct settings are in the table below, but the first number in the scenario name is the number of allies (n_units) and the second is the number of enemies (n_enemies).

Scenario Config File n_units n_enemies
protoss_5_vs_5 sc2_gen_protoss.yaml 5 5
zerg_5_vs_5 sc2_gen_zerg.yaml 5 5
terran_5_vs_5 sc2_gen_terran.yaml 5 5
protoss_10_vs_10 sc2_gen_protoss.yaml 10 10
zerg_10_vs_10 sc2_gen_zerg.yaml 10 10
terran_10_vs_10 sc2_gen_terran.yaml 10 10
protoss_20_vs_20 sc2_gen_protoss.yaml 20 20
zerg_20_vs_20 sc2_gen_zerg.yaml 20 20
terran_20_vs_20 sc2_gen_terran.yaml 20 20
protoss_10_vs_11 sc2_gen_protoss.yaml 10 11
zerg_10_vs_11 sc2_gen_zerg.yaml 10 11
terran_10_vs_11 sc2_gen_terran.yaml 10 11
protoss_20_vs_23 sc2_gen_protoss.yaml 20 23
zerg_20_vs_23 sc2_gen_zerg.yaml 20 23
terran_20_vs_23 sc2_gen_terran.yaml 20 23

Training Results

The smacv2 repo contains the results of MAPPO and QMIX baselines that you can compare now. Please ensure that you are using the correct version of starcraft as otherwise your results will not be comparable. Using the install_sc2.sh in the mappo repo for example will ensure this.

Modifying SMACv2

SMACv2 procedurally generates some content. We encourage everyone to modify and expand upon the procedurally generated content in SMACv2.

Procedurally generated content conceptually has two parts: a distribution and an implementation. The implementation part lives in the starcraft2.py file and should handle actually generating whatever content is required (e.g. the spawning units at the correct start positions) using the StarCraft APIs given a config passed in at the start of the episode to the reset function.

The second part is the distribution. These live in distributions.py and specify the distribution the content is generated according to. For example start positions might be generated randomly across the whole map. The distributions.py file contains a few examples of distributions for the already implemented generated content in SMAC.

Code Example

SMACv2 follows the same API as SMAC and so can be used exactly the same way. As an example, the below code allows individual agents to execute random policies. The config corresponds to the 5 unit Terran map from SMACv2.

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from os import replace

from smacv2.env import StarCraft2Env
import numpy as np
from absl import logging
import time

from smacv2.env.starcraft2.wrapper import StarCraftCapabilityEnvWrapper

logging.set_verbosity(logging.DEBUG)

def main():

    distribution_config = {
        "n_units": 5,
        "n_enemies": 5,
        "team_gen": {
            "dist_type": "weighted_teams",
            "unit_types": ["marine", "marauder", "medivac"],
            "exception_unit_types": ["medivac"],
            "weights": [0.45, 0.45, 0.1],
            "observe": True,
        },
        "start_positions": {
            "dist_type": "surrounded_and_reflect",
            "p": 0.5,
            "n_enemies": 5,
            "map_x": 32,
            "map_y": 32,
        },
    }
    env = StarCraftCapabilityEnvWrapper(
        capability_config=distribution_config,
        map_name="10gen_terran",
        debug=True,
        conic_fov=False,
        obs_own_pos=True,
        use_unit_ranges=True,
        min_attack_range=2,
    )

    env_info = env.get_env_info()

    n_actions = env_info["n_actions"]
    n_agents = env_info["n_agents"]

    n_episodes = 10

    print("Training episodes")
    for e in range(n_episodes):
        env.reset()
        terminated = False
        episode_reward = 0

        while not terminated:
            obs = env.get_obs()
            state = env.get_state()
            # env.render()  # Uncomment for rendering

            actions = []
            for agent_id in range(n_agents):
                avail_actions = env.get_avail_agent_actions(agent_id)
                avail_actions_ind = np.nonzero(avail_actions)[0]
                action = np.random.choice(avail_actions_ind)
                actions.append(action)

            reward, terminated, _ = env.step(actions)
            time.sleep(0.15)
            episode_reward += reward
        print("Total reward in episode {} = {}".format(e, episode_reward))

if __name__ == "__main__":
    main()

Citation

If you use SMACv2 in your work, please cite:

@inproceedings{ellis2023smacv2,
    title={{SMAC}v2: An Improved Benchmark for Cooperative Multi-Agent Reinforcement Learning},
    author={Benjamin Ellis and Jonathan Cook and Skander Moalla and Mikayel Samvelyan and Mingfei Sun and Anuj Mahajan and Jakob Nicolaus Foerster and Shimon Whiteson},
    booktitle={Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track},
    year={2023},
    url={https://openreview.net/forum?id=5OjLGiJW3u}
}

FAQ

Why do SMAC maps not work in SMACv2?

For now, SMAC is not backwards compatible with old SMAC maps, although we will implement this if there is enough demand.

Questions/Comments

If you have any questions or suggestions either raise an issue in this repo or email Ben Ellis and we will try our best to answer your query.