This repository explores the behavioral evolution of NN agents in the prisoner's dilemma. Currently, it implements a two-stage process for playing against opponents. The first stage is an LSTM that is used to predict the strategy of the opponent. The second stage is a set of Q tables that are trained to play against each strategy independently.
To train all models on an agent configuration, run:
python main.py -a <path_to_agent_config> -t -s <save_filename>
To train only the q_table/lstm run:
python main.py -a <path_to_agent_config> -t -s <save_filename> -m qtable
python main.py -a <path_to_agent_config> -t -s <save_filename> -m lstm
To evaluate the performance of the trained models, run:
python main.py -a <path_to_agent_config> -l <load_filename>
To run and visualize a tournament, run:
python main.py -a <path_to_agent_config> -r -n <save_filename> -v
To save visualizations for any command (if possible), add -v to the command.
To create new agent types, new strategies, and new tournament configurations check the Agent Documentation. Note that whenever a new agent type or strategy is added to the AGENT_DICT, the model must be retrained (otherwise results are unknown). Also note that all agent types must have a different ID.
All Model hyper-parameters are kept in the params.py file.