Pokemon_RL_Model

This project is a Reinforcement Learning model designed for Pokémon battles. It uses the poke-env library to interact with the Pokémon environment and implements a Deep Q-Network (DQN) for decision-making. The model is trained to understand Pokémon type effectiveness and can be used to simulate battles and strategies in a Pokémon game environment.
Requirements

1. Have an updated version of Node.js installed
2. Have Python 3.11 installed, or a similar version
3. pip install asyncio
4. pip install tabulate
5. pip install numpy
6. pip install keras==2.12.0
7. pip install poke-env==0.7.0
8. pip install tensorflow==2.12.0 (Mac Users: pip install tensorflow-macos-2.12.0)
9. pip install gym==0.26.2

Usage

1. Clone the pokemon showdown source code repo: https://github.com/smogon/pokemon-showdown
2. Run the command

npm install

This should install all dependencies
3. Run the command

node pokemon-showdown start --no-security

This should create a private showdown server hosted locally. Once the server is running, you can begin running the actual project
4. Clone this repo (Pokemon_RL_Model)
5. Run reinforcement_bot.py

model_filepath = './models/MODEL_NAME.h5'

To use an existing model, change the code above to the appropriate filepath
To start training a new model, set the filepath to a name that doesn't exist already

Models
model.h5:
A simple model trained with 10000 steps. The reward system is as follows:
Winning corresponds to a positive reward of 30
Making an opponent’s pokemon faint corresponds to a positive reward of 1
Making an opponent lose % hp corresponds to a positive reward of % Punishments are mirrored (ex. losing is -30)