IJCAI 2022- Nerual MMO Challenge FAQs

Here is the official FAQ channel, we will update the content continuously.
If you have any issues or something that needs to be discussed, the Discord and QQ support channels will help you.

Q1. Is it possible to use external dependency packages?
A1: All dependency packages must all be installed via mainstream public pip sources and should not contain any non-public binary libraries.

April 14 - April 19 :raised_hand_with_fingers_splayed:

Q1:Could we submit a Script AI?
A1:Sure,you can do anything you want.

Q2:“pip nmmo” could not be found during the installation?
A2:The pip source in China may not be found, changing the source can solve this issue usually. You can try to use the official source (ex: Tsinghua source).

Q3:Could we modify the Reward Parse to increase the rewards for different tasks in the config?
A3:The design of the Reward is a part of the consideration in the challenge. The baseline only provides a simple implementation which would limit the effect. According to your own understanding, participants can modify Reward Parser, feature Parser, and other parts.

Q4:How could I check the effect after training?
A4:In addition using

python -m torchbeast.monobeast --mode test_render --savedir ./results --xpid nmmo

you also need to download the local client(https://github.com/NeuralMMO/client).
The more recommended method is to use the officially provided Rollout for local evaluation; refer to theLocal evaluation.

Q5:Which bots are used in each stage of the PvE evaluation?
A5:Stage1 uses a built-in Scripted AI with 7 Random Teams, 5 ForageTeams, and 3 CombatTeams, while Stage2 and Stage3 use a stronger RL model that has been trained for them.

Q6:How to visualise the replay locally?


–Plan A: You need to aforehand download and open the client for rendering the screen, refer to: Neural MMO Environment Tutorial–Replay
–Plan B:Using terminal viewer contributed by lucienzhu, refer to: Termial Viewer

Q7: What is the baseline algorithm?

April 20 - April 26 :raised_hand_with_fingers_splayed:

Q1:Limited training effect of Baseline?
A1:Supplementary instruction of monobeast baseline:
1) The previous baseline is just an easy implementation of the RL algorithm, which is weaker than the Stage 1 built-in AIs. We also upgraded our training baseline, which could achieve ~0.7 Top1 Ratio in Stage 1 after training for 2 days.
2) Participants could improve from the upgraded baseline to get a stronger AI by paying more attention to feature design, model design or training algorithms.
3) IJCAI 2022-Neural MMO is not limited to RL methods; participants could write their own scripted AI or use other methods to improve their models.

Upgraded baseline: https://gitlab.aicrowd.com/neural-mmo/ijcai2022-nmmo-baselines

Q2:Limitation of training resources?
A2:This challenge does not limit the training resources, and only the resources in the evaluation (inference) stage are limited. Participants can rest assured of training.

April 27 - May 17 :raised_hand_with_fingers_splayed:

Q1:Attack distance calculation method?
A1: The attackable area of agents in NMMO isn’t calculated by L1 distance, is:
the following code could help you figure out the attack calculation method:

Q2:Is Stage2 AI callable during training?
A2:At present, the Stage 2 Neural AI of the competition is closed source and cannot be called during training.

Q3:Is there any critical strike rating in NMMO?
A3:NMMO fixed 5% critical strike rating.


Little bug in Map size?

It seems that the final generated map size is 129*129 instead of 128*128 LOL.

I checked the map file and found that the range that the agent can reach is from (16, 16) to (144, 144).

will you fix this little bug?

Hi, thanks a lot for your reminder. We also comfirmed this bug.

It is in the half of the competition, and there are RL teams using the env to train agents. We do not plan to fix this little bug.

We will fix it in the next NMMO competition.

Best regards,
kirsty, on behalf of IJCAI2022-NeuralMMO team