By Julian Rice, Arman Rafian, Trevor Kirkby, & Joel Valdovinos
A project for CSC 570: AI in Games with Dr. Canaan, at California Polytechnic State University.
This is an agent capable of playing the social deduction game Werewolf, following the specifications of the AIWolf Project. This repository has been forked from the official repository by the AIWolf project, and was originally created by Kei Harada. It also contains a local copy of the AIWolf Server software, which enables running simulated Werewolf games.
This repository includes the source code for a number of previous years' agents for simulation purposes, which can be found on the AIWolf site. The following agents are included in other_agents:
- HALU
- Karma
- OKAMI
- TakedaAgent
- TeamSashimi
- TeamTomato
- TOKU
- TOT Additionally, multiple sample agents provided by the compeitions organizers are included.
A conda environment file is included, which contains the specifications of all python packages needed to run our agent. You can create the environment like the following, if you have conda set up already:
conda env create -n aiwolf -f env.yml
And then to activate:
conda activate aiwolf
There are multiple components to the agent. To run the final product, run_simulations.py is the easiest interface. Run in the following manner:
python3 run_simulations.py --sims N --games 100 --use PolyWolf --name demo
This will run N simulations, of 100 games each, which will use a random selection of opponents and at least one copy of the PolyWolf agent. Logs will be output to sims/15player_100game/demo, and results will be printed to the screen when the sims complete.
To train the neural network yourself on our pre-generated logs, navigate to our_agent/role_estimation and run:
python3 train_rnn.py --lr 0.1
This script will save the best model to model.h5. A pretrained model is provided in pretrained_model.h5, which is what is loaded by the RNN estimator during gameplay. It trained for over 100 epochs. To use your trained model instead, rename it to overwrite pretrained_model.h5.
To generate plots from the paper, navigate to AIWolf-ver0.6.3/rounds_results/. Run python3 graph.py base.txt for voting accuracy plots. Run python3 graph_winrates.py latest_results.txt for win rate plots (the voting accuracy plots from that txt file are innaccurate, because of a now-fixed bug in voting accuracy aggregation).