Game Solving with Online Fine-Tuning

This is the official repository of the NeurIPS 2023 paper Game Solving with Online Fine-Tuning.

If you use this work for research, please consider citing our paper as follows:

@inproceedings{
  wu2023game,
  title={Game Solving with Online Fine-Tuning},
  author={Wu, Ti-Rong and Guei, Hung and Wei, Ting Han and Shih, Chung-Chin and Chin, Jui-Te and Wu, I-Chen},
  booktitle={Thirty-seventh Conference on Neural Information Processing Systems},
  year={2023},
  url={https://openreview.net/forum?id=hN4qpvGzWn}
}

The following instructions are prepared for reproducing the main experiments (baseline, online-cp, online-sp, online-sp+cp) in the NeurIPS paper.

Prerequisites

The game solver program requires a Linux operating system and at least one NVIDIA GPU to operate.

Nevertheless, to reproduce the main experiments in the NeurIPS paper, it is assume that there are three machines, HOST1, HOST2, and HOST3, each of which meets the requirements:

• At least 16 CPU threads and 192G RAM.
• Four NVIDIA GPU cards (GTX 1080 Ti or above).
• Properly installed NVIDIA drivers, podman (or docker), and tmux.

Note that these are not strict requirements of the program. You may run the solver on a single host with only one GPU installed, just use the same machine for HOST1, HOST2, and HOST3 in the instructions below.

Build Programs

Clone this repository with the required submodules:

git clone --recursive --branch neurips2023 git@github.com:rlglab/online-fine-tuning-solver.git
cd online-fine-tuning-solver

Enter the container to build the required executables:

scripts/start-container.sh

# run the below build commands inside the container
scripts/build.sh killallgo # build the main game solver program
cd chat && make && cd .. # build the message service for distributed computing

exit # exit the container

Setup Distributed Computing Nodes

Now we will set up three computing nodes for the distributed game solver. In the instructions below, remember to change HOST1, HOST2, and HOST3 to your actual machine host names.

Launch a new tmux window and start the chat service on HOST1:

scripts/chat/start-service.sh 8888 # use port 8888

Hint: Once the script has been started successfully, it outputs a single line showing the executed command. Just detach the relevant tmux window and let it run in the background.

Deploy the distributed computing nodes on HOST2 and HOST3:

# run this script on both HOST2 and HOST3
scripts/setup-solver-tmux-session.sh CHAT=HOST1:8888

Hint: The script launches a new tmux session named solver. Just simply detach it as previously.

Configure the required workers for either the baseline or the online fine-tuning solver on HOST1 as follows:

export CHAT=HOST1:8888
# run this for the baseline solver:
yes | scripts/run-workers.sh killallgo "8B" HOST2 HOST3
# or, run this for the online fine-tuning solver:
yes | scripts/run-workers.sh killallgo "8O" HOST2 HOST3

Hint: After the successful configuration, the script prints the broker name required for subsequent steps. E.g., b2 has been configured ... indicates a broker named b2.

Solve Opening Problems

Finally, start the container to solve opening problems on HOST1 as follows.

scripts/start-container.sh
# run the below commands inside the container

# to solve JA using the baseline solver (remember to set CHAT and BROKER accordingly)
scripts/run-manager.sh killallgo JA CHAT=HOST1:8888 BROKER=b2 CUDA_VISIBLE_DEVICES=0 TIMEOUT=173000 actor_num_simulation=10000000 use_online_fine_tuning=false

# to solve JA using the online fine-tuning solver (online-cp)
scripts/run-manager.sh killallgo JA CHAT=HOST1:8888 BROKER=b2 CUDA_VISIBLE_DEVICES=0 TIMEOUT=173000 INFO=cp actor_num_simulation=10000000 use_online_fine_tuning=true use_critical_positions=true use_solved_positions=false
# to run online-sp or online-sp+cp solver, set these configs accordingly: INFO, use_critical_positions, and use_solved_positions

Hint: To solve the opening using the baseline (or online fine-tuning) solver, the computing nodes must be configured accordingly by running run-workers.sh with 8B (or 8O) before running run-manager.sh.

Results are stored after a successful run. For example, the results of solving JA with the online fine-tuning solver (where $NAME is JA-online-10000000-16.5-gpcn_n32_m16_empty_op-100000-7g384w+1sp1op-cp):

• $NAME.log stores the log output by the manager.
• $NAME.sgf stores the search tree. Use GoGui to open the file.
• $NAME.stat stores the main result and statistics. Use scripts/extract-main-stat.sh $NAME.stat to extract fields as shown in the paper.
• $NAME.jobs stores all jobs generated by the manager.
• training/online_$NAME stores the network models generated during the online fine-tuning.

Detailed Instructions

1. Development Environment covers the basic of developing with containers and building programs.
2. Train PCN Models for the Solver introduces on the training of Proof Cost Networks.
3. Launch the Standalone Solver instructs how to solve problems using the worker.
4. Launch the Distributed Solver instructs how to solve problems using the manager, the workers, and the learner.
5. Tools describes the usage of miscellaneous tools related to the game solvers.

References

• Game Solving with Online Fine-Tuning
• A Novel Approach to Solving Goal-Achieving Problems for Board Games
• AlphaZero-based Proof Cost Network to Aid Game Solving
• A Local-Pattern Related Look-Up Table