final solution

README.md

@@ -1,10 +1,48 @@
-![figure](https://neuralmmo.github.io/_static/banner.png)
+# The Winning Solution for the NeurIPS 2023 Neural MMO Challenge
 
-# ![icon](https://neuralmmo.github.io/_build/html/_images/icon.png) Welcome to the Platform!
+This solution is based on the [Neural MMO Baselines](https://github.com/NeuralMMO/baselines/tree/2.0?tab=readme-ov-file). For more information about the challenge, please refer to the [challenge homepage](https://www.aicrowd.com/challenges/neurips-2023-the-neural-mmo-challenge).
 
-[![](https://dcbadge.vercel.app/api/server/BkMmFUC?style=plastic)](https://discord.gg/BkMmFUC)
-[![Twitter](https://img.shields.io/twitter/url/https/twitter.com/cloudposse.svg?style=social&label=Follow%20%40jsuarez5341)](https://twitter.com/jsuarez5341)
+## How to Use
 
-Baselines repository for Neural MMO. New users should treat this as a starter project. The project is under active development with [Documentation](https://neuralmmo.github.io "Neural MMO Documentation") hosted by github.io.
+### Environment Installation
+Use `docker/build.sh` to build image for training.
+```shell
+cd docker
+bash build.sh
+```
 
-![figure](https://neuralmmo.github.io/_build/html/_images/poster.png)
+### Run training
+Run inside the training container:
+```shell
+export WANDB_API_KEY=xxx  # Change it to yours
+WANDB_PROJECT=xxx  # Change it to yours
+WANDB_ENTITY=xxx  # Change it to yours
+
+export WANDB_DISABLE_GIT=true
+export WANDB_DISABLE_CODE=true
+
+export OMP_NUM_THREADS=4
+
+python train.py \
+       --runs-dir runs \
+       --use-ray-vecenv true \
+       --wandb-project $WANDB_PROJECT \
+       --wandb-entity $WANDB_ENTITY \
+       --model ReducedModelV2 \
+       --meander-bonus-weight 0.0 \
+       --heal-bonus-weight 0.0 \
+       --num-npcs 128 \
+       --early-stop-agent-num 0 \
+       --resilient-population 0.0 \
+       --ppo-update-epochs 1 \
+       --train-num-steps 40000000 \
+       --num-maps 1280 \
+```
+
+### Evaluation
+
+After training, copy the checkpoints into `policies` and run:
+```shell
+python evaluate.py -p policies
+```
+`policies/submission.pkl` is the trained model we submitted.

docker/Dockerfile

@@ -0,0 +1,10 @@
+ARG DOCKER_BASE_IMAGE
+FROM ${DOCKER_BASE_IMAGE}
+
+COPY requirements.txt ./
+
+RUN sudo rm /etc/apt/sources.list.d/* && sudo apt-get update && sudo apt-get install -y vim
+
+RUN $HOME/anaconda3/bin/pip --no-cache-dir install -r requirements.txt
+
+RUN sudo rm requirements.txt

docker/build.sh

@@ -0,0 +1,5 @@
+DOCKER_BASE_IMAGE="${DOCKER_BASE_IMAGE:-rayproject/ray:1.6.0-py39-gpu}"
+
+docker build --no-cache \
+--build-arg DOCKER_BASE_IMAGE=${DOCKER_BASE_IMAGE} \
+-t ${DOCKER_BASE_IMAGE}-nmmo2023 .

requirements.txt → docker/requirements.txt

@@ -9,6 +9,10 @@ scikit-learn==1.3.0
 tensorboard==2.11.2
 tiktoken==0.4.0
 torch==1.13.1
+torchtyping==0.1.4
 traitlets==5.9.0
 transformers==4.31.0
 wandb==0.13.7
+
+nmmo==2.0.3
+pufferlib==0.4.5