Learn to play Snake using Reinforcement Learning

Overview

The concept is as follows: the player maneuvers a line which grows in length, with the line itself being a primary obstacle. The line resembles a moving snake that becomes longer as it eats food. The player loses when the snake runs into the screen border or itself. The goal is to get as long as possible before dying. In this project, we introduce our different reinforcement learning approaches in order to efficiently train an agent to play Snake.

Quick Start

Install all required packages:

pip3 install -r requirements.txt

In main.py choose the hyper-parameters, the model and run:

python3 main.py

Snake is learning!

Train Models

4 RL models are currently implemented in the framework. 3 classic models and 1 deep learning model. All models inherit from the ABCTrainer class which contains all the abstract methods needed to create a new model (choose_action, update_hyperparameters...). Check in the trainer folder what the hyperparameters needed are for each models.

Play With A Pretrained Model

• For classic models: in main import a pretrained q_table by using the import_q_table function to instanciate the trainer.
• For DQLTrainer: use the load_model_path attribute of DQLTrainer, we provide pytorch pretrained weights that you can use to save you hours of training before getting results.
• Be careful to put epsilon_init = 0 when using a pretained model.

epsilon_init = 0
DQLTrainer(epsilon_init, epsilon_decay, learning_rate, gamma, decay_rate, 
    size_x, size_y, load_model_path="pretrained_models/trained_weights.pth")

It is also possible to stop the training of your model, by using the test argument of the trainers' iterate method.

Find the best hyper-parameters

grid_search.py allows you to test a wide range of hyper-parameters on the model of your choice, it returns a json file with the results.

More details

More details can be found in our paper Snake-RL.pdf.