PlotMap: Automated Layout Design for Building Game Worlds [paper]

Environment Setup

Tested under Python 3.9. Dependent on the following libraries:

• Pytorch 2.0.x
• Ray(RLLib) 2.0.1
• Shapely 1.8.4
• Gym 0.23.1
• cma 3.4.0
• numpy 1.23.1
• opencv-python 4.10.0.84
• img2vec_pytorch 1.0.1
• sentence_transformer 3.0.1
• If there are other missing libraries, they can be simply installed with pip install

Data

Task sets can be downloaded from https://huggingface.co/datasets/ywang485/plotmap-taskset

CMA-ES solving/inference

Commands for running CMA-ES solving on example task

python fixed_solve.py --task_file='tasksets/task_1.json' --task_name='test_task' --output_file='results.json' --max_fevals=500 --fixed_indices=[0,1,2,3] --fixed_values=[35,95,6,15]

RL Single Rollout

Use the following command to execute a single rollout

run_agent.py path_to_checkpont path_to_task

If successful, a json file namned acility_positions.json will be created in the current work directory with rollout information.

RL Model Training

Modify facility_placement_ppo.yaml as needed, then execute

python train.py -f facility_placement_ppo.yaml

RL inference

After setting the path to the task set and the model to test with in line 22-23 in run_gym.py, execute

python run_gym.py

1000 rollouts will be performed and success rate will be reported after loading the task set (which may take some time).