Snake Game 🐍

This final "Reinforcement Learning course" project focuses on the artificial intelligence of the "Snake Game". Essence of the game is a piece of cake: snake eats the food that arises on the field and fill the map of its bodies as soon as possible.

We have implemented three algorithms:

• Value Iteration
• Policy Iteration
• Simple Multi-Layer perceptron

In order to estimate the perfomance of each algorithm we have conducted an experiments, using the following metrics:

• Average number of steps (ANS): How many steps does the Snake take on average before the collision;
• Average maximal Snake length (AML): Average maximal length of the snake during the game;

Our video-presentation can be found here.

Results

Algorithm	Value Iteration	Policy Iteration	MLP
Demo (Non-optimal)
Average number of steps	424	198	123
Average maximal length	22	23	12

Repository structure

• models/ — contains algorithms
  ├── nn.py — MLP approach
  ├── policy_iteration.py — Policy Iteration algorithm
  └── value_iteration.py — Value Iteration
• src/ — auxiliary instrumets
  ├── render.py — Snake Game rendering
  └── snake_game.py — logic of the Snake Game
• gifs/ — contains gifs
• supplementary/ — contains presentation in .pdf format
• run.py — start experiments

Installation

Follow the rows:

$ git clone https://github.com/queen-s0/snake_game.git
$ cd snake_game
$ conda env create -f environment.yml

Usage

usage: run.py [-h] [--algo {value iteration,policy iteration,mlp}]
              [--field-shape FIELD_SHAPE [FIELD_SHAPE ...]]
              [--train-games TRAIN_GAMES] [--visualize VISUALIZE]
              [--seed SEED] [--test-games TEST_GAMES]

Example

python run.py --algo 'mlp' --field-shape 15 15 --train-games 8000 --visualize True --test-games 5