Code for my experiments with training a battle agent for the original Pokemon Red. Read the report here.
After getting the source code, grab all the necessary requirements for the project by running the command pip install -r requirements.txt in the local project directory.
The emulator is provided by the useful module pyboy.
The wrapper code for Pokemon Red is written by me, and located in battle_wrapper.py. The demo for this wrapper is located at emu_demo.py. There are some variables at the top of the file that you may change. To fully control the emulator with arrow keys, X, and Z, set DEMO_MODE to false. There are two demo modes, a manual stepper and a random stepper. Change which one to use by editing MANUAL_STEP_MODE. Be advised, the manual stepper will ask for user input.
Any files for game ROMs, RAMs, and states are located in the /roms directory.
The code for the environment, which mirrors that of an OpenAI gym environment, can be found in pokemon_env.py. The code for the DQN agent, including the neural network, replay buffer, DQN implementation, and training functions is located in dqn.py.
To train a model, run the script train.py. Note that any changes to the network parameters and dimensions must also be reflected in the environment you choose to use. For example, for my two experiments, I have two seperate environments - PokemonEnvironment and UpdatedPokemonEnvironment (aptly named).
The model training function will output three numpy arrays and multiple PyTorch network files along the way, including temporary, final, and best network (which generated the highest return in an episode).
To demo my experiments, run the script dqn_demo.py. There are a few options available which can be toggled by changing some variables. When enabling TRIAL_MODE, a trial of 100 runs of the battle will be conducted at the fastest possible speed (the same speed at which the model was trained). Data for total wins, losses, and returns will be output at the end of the trial. MODEL_NUM lets you specify which of my models will be tested, 1 or 2.
The results from my two experiments are located in the notebook file, run_experiments_plots.ipynb.
Any data generated from my experiments is located in the /data folder.