Javascript must be enabled for the correct page display

Reinforcement learning for the game of soccer in flexible environments

Hegeman, Julian and Kim, Yujin (2019) Reinforcement learning for the game of soccer in flexible environments. Bachelor's Thesis, Artificial Intelligence.


Download (307kB) | Preview
[img] Text
Restricted to Registered users only

Download (140kB)


The soccer game is one of the most popular sport games. Even though the game is easy to play for humans, it requires a wide range of intelligent abilities. With this lens, the game of soccer is a very intriguing topic in Artificial Intelligence. In this research, we apply reinforcement learning techniques, which have been successful in various games, such as Go and Atari. Using these techniques, we focus on whether it is possible to design a system that performs in simulations with multiple team and field sizes at an equal level as a multilayer perceptron (MLP) based Q-learning agent that receives information from all points in a grid modeling the game of soccer and is trained specifically on one field size. This reference player is also trained on varying team sizes. The results illustrate that the vision grid player performs better than the reference player on all tested setups except the one with the smallest team and field size (9 by 9 cells and 1 player per team). The vision grid player is therefore flexible regarding the size of the field it plays in and the number of players it plays with, and able to outperform the reference player that is only trained using a varying number of players per team. The influence of the activation function in the hidden layer of the MLP is also researched. It is found that for this simulation, the sigmoid activation function works best.

Item Type: Thesis (Bachelor's Thesis)
Supervisor name: Wiering, M.A.
Degree programme: Artificial Intelligence
Thesis type: Bachelor's Thesis
Language: English
Date Deposited: 25 Sep 2019
Last Modified: 26 Sep 2019 11:09

Actions (login required)

View Item View Item