YuhangSong
commented
4 years ago Hi,
Sorry for the late reply, I have been traveling due to the coronavirus.
The baseline tutorial videos seems still private, and cannot be played;
Please simply follow the README instructions in this repo, these videos have been depreciated so not useful anymore. The README instructions have everything you need.
In the /arena-experiments/Arena-Benchmark.yaml file, there are four env examples at the top, like "Arena-Tennis-Sparse-2T1P-Discrete". What other envs do you have; what the corresponding names are; any document specifies this?
We are re-benchmarking all games, currently, all games are available in Arena-BuildingToolkit. We are building and testing them one by one. You can build and test it yourself if you urgently want to try some the games (we have played with the games and they are fine, it's just we have not have enough resources to train them yet). Before you build a game and train it, you can also play with it in Arena-BuildingToolkit so that you can make sure you are satisfied with the game. The four env examples you are seeing are the four ones we have built and trained.
The /arena-experiments/Arena-Benchmark.yaml file is an example using PPO, "run: PPO", how can I use other baseline methods, what are the corresponding names of other methods, such as COMA or MADDPG? Where are the scripts of those methods?
COMA should be easy to have with the current config of "Use separated observations for actor and critic (such as centralized critic and decentralized actors)." For MADDPG, we should be able to integrate it from rllib. We are working on these two baselines and will keep you updated.
Any instructions about how to test my own algorithms using your envs?
If you are thinking about using a completely different code base, just using our environments, the python file test_arena_rllib_env.py is what you are looking for, it tells you how to use the python interface to interact with the built games.
If you are thinking about code your algorism within our codebase, you should start from reading train.py as well as going through rllib
Please feel free to raise further questions!
Hi,
I am trying to use Arena in my research project. I have several general questions:
1) The baseline tutorial videos seems still private, and cannot be played; 2) In the /arena-experiments/Arena-Benchmark.yaml file, there are four env examples at the top, like "Arena-Tennis-Sparse-2T1P-Discrete". What other envs do you have; what the corresponding names are; any document specifies this? 3) The /arena-experiments/Arena-Benchmark.yaml file is an example using PPO, "run: PPO", how can I use other baseline methods, what are the corresponding names of other methods, such as COMA or MADDPG? Where are the scripts of those methods? 4) Any instructions about how to test my own algorithms using your envs?
I may miss something when I went through the READM.md. If you could point them out, it would be really appreciated!
Best, Yuchen