The Implementation of Reinforcement Learning Algorithm for AI Bot in Fighting Video Game

A video game is a natural and valuable medium to test AI algorithms since this virtual environment is considered safe and controllable. The finite state machine is a common technique used in developing AI in games, using a predetermined action for each situation. The problem with a predetermined action is that the scripted behaviours are predictable and easily exploit by human players. Such settings cause the game to have repetitive gameplay, leading the player to lose interest since the player knows how the AI will behave. The primary approach of this project is to use reinforcement learning to train the agent for the game. Reinforcement learning algorithms learn what to do and map the situations to maximize the cumulative reward signal from the agent environment. The agent in this project will receive raw data from the environment as input, and the agent will then act based on the environment rather than predetermined action. The results show that lowering the learning rate in deep reinforcement learning can increase the cumulative reward for the agent. This project's findings will be helpful for the game developers in developing AI for their games and benefit the game players who will interact with the AI in those games.

I. Introduction

The video game is interactive digital entertainment that people play through multiple medium such as a computer, a game console, or a smartphone. There are different genres of video games, ranging from complex strategy games to simple arcade games. It is stated in a book titled “Artificial Intelligence and Games” that video games are an ideal testbed for artificial intelligence (AI) methods and are also becoming a growingly crucial application area [1]. Games from different genres could provide various problems for agents to solve, making video games an excellent environment for AI research. Not only that, but these virtual environments are also actually safe and controllable [2].