Skip to content

Latest commit

 

History

History
49 lines (38 loc) · 2.87 KB

README.md

File metadata and controls

49 lines (38 loc) · 2.87 KB
Raw

Modern Game AI Algorithms Project: PCG and Kiting in 0 A.D.

In order to run the project, you will need to install version 0.24 of 0 A.D., available here.

You can add pyrogenesis to your environment variables for easier use, otherwise always specify the path to the execution file.
In Windows, that file should be: game_install_path/binaries/system/pyrogenesis.exe
In MacOS, the file should be: /Applications/0 A.D..app/Contents/MacOS/pyrogenesis

You will also need to add the RL-scenarios mod to 0 A.D., which we have provided in the zero_ad_rl folder (zipped folder 0ad_rl_maps.zip). To add the mod to 0 A.D., run pyrogenesis /path/to/mod/mod.zip in a terminal.

Project Setup

We recommend you create a new conda environment to run the project.
To install the project's required packages, you will need to run:

python install zero_ad/.
python install -e zero_ad_rl/.
python install -r requirements.txt

Training Agent

Game Launch

To train an agent in a specific environment, you will first need to start 0 A.D. When training an agent, we suggest that you start 0 A.D. with the graphics disabled as this will speed up the training process. This can be done by running the following command in a terminal:

pyrogenesis --rl-interface=0.0.0.0:6000 --autostart-nonvisual --mod=rl-scenarios --mod=public

As you can see, you also need to specify the mods to launch and use, as well as assign a port for the python scripts to connect to.

Training

To train an agent, you can then use the following command:

python -m zero_ad_rl.train --env CavalryVsInfantry --run PPO --checkpoint-freq 25

You can change the training algorithm with --run <algo> by specifying one of rllib's built-in algorithms. You can also switch which environment to train the agent on using --env <env_name>. The options are; CavalryVsInfantry, CavalryVsInfantryMaze and CavalryVsInfantryCity. The code uses rllib to train agents register on a defined environment.

Rollout

Finally, to perform rollout on the trained agent, we recommend that you stop the current running process of 0 A.D. to run it with graphics enabled (simply remove --autostart-nonvisual).
You can then run the rollout function on a specific training checkpoint with the following command:

python -m zero_ad_rl.rollout_pcg ~/ray_results/path/to/checkpoint/file --env CavalryVsInfantry --run PPO --pcg --steps 5000

By default, the rollout will not generate a new map to evaluate the agent on, simply add/remove --pcg to turn this option on/off.

Note: an agent can have been trained on an evironment (say CavalryVsInfantry) and evaluated using a different one (say CavalryVsInfantryMaze). Simply specify the CavalryVsInfantryMaze enviroment when performing the rollout.