AI Learns to play Tetris by evolution

AE4350 Bio-inspired Intelligence and Learning

This repository contains the code used for the AE4350 assigment. In this assignment, an AI is evolved to play Tetris by use of the gym-tetris environemt and pytorch.

Prerequisites

• python 3.8
• cudatoolkit 11.6
• pytorch 1.12 (cuda 11.6 version)
• gym
• gym-tetris

Note that the code can also be used without a dedicated GPU, this required changes to the code however. The included requirements.txt can be used to create a suitable conda enviromnent by entering the following into an Anaconda prompt:

conda create --name <env> --file requirements.txt

with <env> the name you want to give the environment.

Perform evolution

Here's an overview of how the evolution works:
In run_evolution.py the following evolution parameters can be set:

Parameter	Description
run	run ID
net_type	network type (non-CNN, CNN)
net_size	network size
generations	number of generations
individuals	number of individuals per generation
indiv_best	number of best individuals used for mutation
determ_bool	determinism of tetris game
prob_mut	mutation probatility
p_mut	mutation factor
p_prep	playfield preparation factor for tertary representation
p_height_reward	reward factor for height reward
p_edge_reward	reward factor for edgereward
p_rwd2	reward double if piece tetromino falls perfectly
p_rwd_NOOP	reward factor for no-operation action

When running an evolution, a folder ./runs/run<ID> is created where the generation tensors and rewards are stored.

Evaluate an evolution

The performance of a completed evolution can be evaluated by use of one of the eval python scripts (use the applicable one based on network type). These generate a plot of the rewards vs generations and render a game with the best performing performing individual.