First, clone this directory and its submodules (necessary for only for evolving diverse generators):
git clone --recurse-submodules https://github.com/smearle/control-pcgrl
cd control-pcgrl
(repo address will be [email protected]:smearle/control-pcgrl if using SSH).
Then, (un)comment the relevant lines in setup.sh to install torch with(out) GPU/CUDA enabled for your machine. Then run:
bash setup.sh
Note that the pytorch-neat and qdpy submodules are only necessary for evolving diverse generators (not RL).
Run experiments by editing the batch.yaml files in either configs/evo or configs/rl, then running run_batch_evo.py or run_batch_rl.py, which in turn launch the scripts evo/evolve.py or rl/train_ctrl.py, respectively, either locally (sequentially, with option --local), or on SLURM (in parallel).
See gym_pcgrl for the original Readme, from amidos2006/gym-pcgrl.
The below instructions, for controllable RL and evolving diverse generators, respectively, are out of date.
Clone this repository along with its submodules:
git clone --recurse-submodules -j8 https://github.com/smearle/control-pcgrl
It is recommended to use a virtual environment using anaconda or similar.
conda create -n pcgrl python=3.10
conda activate pcgrl
To install the required python packages, it should suffice to run
python -m pip install -r requirements.txt
If the above causes errors, the offending lines can be commented out, and any missing packages installed manually.
GTK and PyGObject are required for rendering controllable PCGRL agents, but are not used in the evolution pipeline. They can be installed with:
conda install -c conda-forge pygobject gtk3
We use Evocraft to render in Minecraft. Launch an evocraft server (following the instructions at the link) then run inference a trained agent with render=True. TODO: Add ability to launch server from this repo.
Config files and hyperparemeter sweeps can be found in configs/rl.
The command
train_pcgrl
will train an RL generator with the default settings given in config.yaml. (The main config class ControlPCGRLConfig can be found in config.py.)
train_pcgrl -m +experiment=learning_rates
will conduct a hyperparameter sweep over learning rates as defined in learning_rates.yaml
This multirun will launch on SLURM by default. This can be changed in config.yaml or by adding +hydra.launcher=submitit_local to the command line.
This repo contains the code for the paper.
This paper should be cited if 3D code from this project is used in any way:
@inproceedings{jiang2022learning,
title={Learning Controllable 3D Level Generators},
author={Jiang, Zehua and Earle, Sam and Green, Michael and Togelius, Julian},
booktitle={Proceedings of the 17th International Conference on the Foundations of Digital Games},
pages={1--9},
year={2022}
}
Note: this instruction might be somehow outdated. Please check the more up-to-date ones above
This is the codebase used to generate the results presented in the paper available on arxiv. It builds on the codebase for PCGRL, whose readme is included below for reference.
To generate the results presented in the maze domain, in Table 1 of the paper, run python evo_batch.py on a SLURM cluster, or python evo_batch.py --local on a local machine. This will launch a series of experiments (either on separate nodes of a cluster or in sequence on a local machine). If you're on a SLURM cluster, you'll need to replace my email with your own, in evo_train.sh and evo_eval.sh.
The evo_batch.py file essentially repeatedly calls python evolve.py with particular sets of hyperparameters, so you may also want to experiment with calling that file directly---just be sure to take a look at all the arguments (visible with python evolve.py -h or in evo_args.py that can be provided (and note that in the paper, we always provide the flag --fix_elites since the re-evaluation of elites during training was found to mostly have a negative effect on the final archive). Results will be saved in the evo_runs directory, every --save_interval-many generations. Adding the --render option when calling evo_batch.py or evolve.py will render the level-generation process in real-time, during either training or evaluation. The former can be useful for debugging, or to get an intuitive sense of what the system is doing, but note that it will slow training down quite drastically.
To evaluate saved archives, run python evo_batch.py --evaluate (which essentially calls python evolve.py --infer --evaluate). To visualize the results of cross-evaluation in a table, run python evo_batch.py --local --cross_eval --tex (running without --tex will generate a large html with all results instead a of a tex file with only a focused subset of results). The table-generation is automated in evo_cross_eval.py. To render gifs from the level frames that were saved during evaluation, run python evo_batch.py --local --gifs.
This paper should be cited if evo code from this project is used in any way:
@inproceedings{earle2022illuminating,
title={Illuminating diverse neural cellular automata for level generation},
author={Earle, Sam and Snider, Justin and Fontaine, Matthew C and Nikolaidis, Stefanos and Togelius, Julian},
booktitle={Proceedings of the Genetic and Evolutionary Computation Conference},
pages={68--76},
year={2022}
}
This repo also contains the code for the paper presented at CoG 2021.
This code requires pygobject and gtk3 (installation described above), and stable-baselines 2 and tensorflow 1 (installation described below).
To train, visualize, and evaluate a controllable generator, run train_ctrl.py, infer_ctrl.py and evaluate_ctrl.py, respectively.
This paper should be cited if controllable code from this project is used in any way:
@inproceedings{earle2021learning,
title={Learning controllable content generators},
author={Earle, Sam and Edwards, Maria and Khalifa, Ahmed and Bontrager, Philip and Togelius, Julian},
booktitle={2021 IEEE Conference on Games (CoG)},
pages={1--9},
year={2021},
organization={IEEE}
}
python generate_trajectories.py
python train_imitation.pyNOTE: This is out of date and should be updated to reflect changes to the environment.
Current Framework Version: 0.4.0
An OpenAI GYM environment for Procedural Content Generation via Reinforcement Learning (PCGRL).
The framework, along with some initial reinforcement learning results, is covered in the paper PCGRL: Procedural Content Generation via Reinforcement Learning. This paper should be cited if code from this project is used in any way:
@misc{khalifa2020pcgrl,
title={PCGRL: Procedural Content Generation via Reinforcement Learning},
author={Ahmed Khalifa and Philip Bontrager and Sam Earle and Julian Togelius},
year={2020},
eprint={2001.09212},
archivePrefix={arXiv},
primaryClass={cs.LG}
}
- Clone this repo to your local machine.
- To install the package, run
pip install -e .from inside the repo folder. (Don't worry it will install OpenAI GYM environment automatically, otherwise you can install it first by following that link) - If everything went fine, the PCGRL gym interface is ready to be used. Check the following section on how to use it.
The PCGRL GYM interface has multiple different environments, where each environment consists of two parts: a problem and a representation. All the environments follow the following name conventions:
[problem_name]-[representation_name]-[version]
For the full list of supported problems names check the Supported Problems section and for the full list of the supported representations name check the Supported Representations section.
To list all the registered environments, you can run the following code:
from gym import envs
import gym_pcgrl
[env.id for env in envs.registry.all() if "gym_pcgrl" in env.entry_point]After installing the interface, you can use it like any other GYM interface. Here is a simple example on how to use the framework on the Sokoban environment with Narrow representation:
import gymnasium as gym
import gym_pcgrl
env = gym.make('sokoban-narrow-v0')
obs = env.reset()
for t in range(1000):
action = env.action_space.sample()
obs, reward, done, info = env.step(action)
env.render('human')
if done:
print("Episode finished after {} timesteps".format(t+1))
breakBeside the OpenAI GYM traditional functions. Our interface supports additional functionalities such as:
self.get_num_tiles(): This function get the number of different tiles that can appear in the observation spaceget_border_tile(): This function get the tile index to be used for padding a certain problem. It is used by certain wrappers.adjust_param(cfg): This function that helps adjust the problem and/or representation parameters such as modifyingwidthandheightof the generated map.
Problems are the current games that we want to apply PCGRL towards them. The following table lists all the supported problems in the interface:
| Name | Goal | Tile Values |
|---|---|---|
| binary | Generate a fully connected top down layout where the increase in the longest path is greater than a certain threshold | 0: empty, 1: solid |
| ddave | Generate a fully connected level for a simple platformer similar to Dangerous Dave where the player has to jump at least 2 times to finish | 0: empty, 1: solid, 2: player, 3: exit, 4: diamonds, 5: trophy (act like a key for the exit), 6: spikes |
| mdungeon | Generate a fully connected level for a simple dungeon crawler similar to MiniDungeons 1 where the player has to kill 50% of enemies before done | 0: empty, 1: solid, 2: player (max of 5 health), 3: exit, 4: potion (restores 2 health), 5: treasure, 6: goblin (deals 1 damage), 7: ogre (deals 2 damage) |
| sokoban | Generate a fully connected Sokoban level that can be solved | 0: empty, 1: solid, 2: player, 3: crate (to be pushed toward the target), 4: target (the location where the crate should ends) |
| zelda | Generate a fully connected GVGAI zelda level where the player can reach key then the door | 0: empty, 1: solid, 2: player, 3: key (to be collected before the door), 4: door (to win the level), 5: bat (should be avoided), 6: scorpion (should be avoided), 7: spider (should be avoided) |
| smb | Generate a simplified and playable Super Mario Bros level where there is at least 20 jumps in the level | 0: empty, 1: solid, 2: enemy, 3: brick, 4: question, 5: coin, 6: tube (need 2 beside each other) |
Representations are the way the Procedural Content Generation problem is formatted as a Markov Decision Process to be able to use it for reinforcement learning. All the problems can be represented using any of the supported representations. The following table shows all the supported representations in the interface:
| Name | Observation Space | Action Space |
|---|---|---|
| narrow | 2D Box of integers that represent the map and 1D Box of integers that represents the x, y position | One Discrete space that represents the new tile value and no change action |
| narrowcast | 2D Box of integers that represent the map and 1D Box of integers that represents the x, y position | Two Discrete spaces that represent the type of change (no change, single, 3x3 grid) and the new tile value |
| narrowmulti | 2D Box of integers that represent the map and 1D Box of integers that represents the x, y position | Nine Discrete spaces that represent the new tile value and no change action |
| wide | 2D Box of integers that represent the map | Three Discrete spaces that represent the x position, y position, new tile value |
| turtle | 2D Box of integers that represent the map and 1D Box of integers that represents the x, y position | One Discrete space where the first 4 actions move the turtle (left, right, up, or down) while the rest of actions are for the tile value |
| turtlecast | 2D Box of integers that represent the map and 1D Box of integers that represents the x, y position | Two Discrete spaces that represents movement+type and tile values. The first space represents 4 actions to move the turtle (left, right, up, or down) while the rest of actions are type of change (1 tile, 3x3 grid) |
The narrow, wide, and turtle representation are adapted from Tree Search vs Optimization Approaches for Map Generation work by Bhaumik et al.
Create the new problem class in the gym_pcgrl.envs.probs and extends Problem class from gym_pcgrl.envs.probs.problem. This class has to implement the following functions.
def __init__(self):
super().__init__()
...
def get_tile_types(self):
...
def get_stats(self, map):
...
def get_reward(self, new_stats, old_stats):
...
def get_episode_over(self, new_stats, old_stats):
...
def get_debug_info(self, new_stats, old_stats):
...Also, you need to make sure that you setup the following parameters in the constructor:
self._width: the generated map width.self._height: the generated map height.self._prob: a dictionary for all the game tiles where keys are the tile names and the values are the probability of the tile appearing when initializing a random map.self._border_size: the size of the border added around the generated level (in a lot of games there might be a border surrounding the level, it is a good idea to get that out).self._border_tile: the tile name used for the border.self._tile_size: the size of the tile in pixels to be used in rendering.self._graphics: a dictionary for all the game graphics where keys are the tile names and values are the Pillow images for rendering the problem.
Feel free to override any other function if you need a behavior different from the normal behavior. For example: In all our problems, we want our system to not load the graphics unless it is going to render it. We override render() function so we can initialize self._graphics at the beginning of the render() instead of the constructor.
After implementing your own class, you need to add the name and the class in gym_pcgrl.envs.probs.PROBLEMS dictionary that can be found in __init__.py the key name is used as the problem name for the environment and the value is to refer to the main class that it need to construct for that problem.
Create the new representation class in the gym_pcgrl.envs.reps and extends Representation class from gym_pcgrl.envs.reps.representation. This class has to implement the following functions.
def __init__(self, width, height, prob):
super().__init__(width, height, prob)
...
def get_action_space(self):
...
def get_observation_space(self):
...
def get_observation(self):
...
def update(self, action):
...
# boolean to define where the change happened and x,y for the location of change if it happened
return change, x, yFeel free to override any other function if you need a behavior different from the normal behavior. For example: in the narrow representation, we wanted to show the location where the agent should change on the rendered image. We override the render() function to draw a red square around the correct tile.
After implementing your own class, you need to add the name and the class in gym_pcgrl.envs.reps.REPRESENTATIONS dictionary that can be found in __init__.py the key name is used as the representation name for the environment and the value is to refer to the main class that it need to construct for that representation.
train.py uses stable baseline PPO2 algorithm for training. You can configure train.py to train for different problems or different representation by changing the game and representation parameters in the file to a different problem and/or representation.
To read more about the experiments and the different wrappers, check our paper PCGRL: Procedural Content Generation via Reinforcement Learning.
You can run the code either using the Dockerfile using the following command line after installing Docker:
docker image build -t pcgrl:latest . && docker run --runtime=nvidia pcgrl:latestAnother way is to use Conda by creating a virtual environment then activating it and installing all the dependencies for train.py:
conda create --name pcgrl
conda activate pcgrl
pip install tensorflow==1.15
pip install stable-baselines==2.9.0
cd gym_pcgrl
pip install -e .
cd ..
python train.pyLastly, you can just install directly without using any virtual environment:
pip install tensorflow==1.15
pip install stable-baselines==2.9.0
cd gym_pcgrl
pip install -e .
cd ..
python train.pyPS: The training process will create a folder named runs/ where two folders will appear one for tensorboard logs and the other for the saved models. The training is always saving the best model so far and the last model.
First make sure you have all required modules (GYM PCGRL, Tensorflow, and Stable Baselines) are installed either using Docker, Conda, or Pip directly. The code to run is similar to the above code for training just change train.py to inference.py.
In the case, that you want to use jupyter notebook, please check inference.ipynb. Please, make sure to choose the correct kernel (especially if you are using Conda virtual environments) before running anything.
Here is a cool GIF when running these models:
PS: All the models for Sokoban Narrow, Sokoban Turtle, and the third model of Zelda Turtle has been saved using python 3.5 which have a different serialization method than python 3.6 and 3.7. When try to load them in python 3.6 or 3.7, you will get an unknown op code error so make sure that you are using the correct python version. We apologize for this mess and we are working on training new models using python 3.7 to replace these ones. Remember if you get unknown opcode, it is because of the serialization method. We didn't know about that issue until later, sorry again for any inconvenience. One last note, Python 3.6 can't load Binary Narrow and Zelda Narrow so make sure to use python 3.7 for all the models except the one that needs python 3.5.
Bug reports and pull requests are welcome on GitHub at https://github.com/amidos2006/gym-pcgrl.
This code is available as open source under the terms of the MIT License.