Mastering Games with AI

21 Aug 2018

This is an essay I wrote last spring on the use of AI for playing board games. It describes the history behind the AI systems that eventually defeated human experts at playing chess, checkers, backgammon, and Go. The essay was an assignment for a course at Brown called “Reintegrating AI”, and I submitted it on March 23, 2018.

Read the full essay here.


Introduction

Since its inception, the field of artificial intelligence (AI) has sought to measure the intelligence of machines by comparing their performance against that of humans. In Alan Turing’s seminal paper, he suggested that a good way to evaluate machine intelligence is to have machines play games against human adversaries. Intuitively, the argument is that if greater intelligence causes one to be better at winning games, then for a machine to win a game against a human, it must have greater intelligence (at least at playing that game). This conclusion seems especially valid for two-player, zero-sum games, where one player wins if and only if the other loses. A great deal of AI research has thus focused on building AI systems that can play two-player, zero-sum games at a high skill level.

AI approaches to game-playing can largely be grouped into two categories: classical techniques (also known as Good Old-Fashioned AI, or GOFAI), and connectionism. Classical techniques tend to focus on explicitly representing abstract knowledge about the game in some sort of knowledge base, and then efficiently searching through that knowledge when it’s time to select the next move. In connectionist approaches, the system’s knowledge is represented implicitly within one or more neural networks, and these networks process the game state as input before outputting an evaluation of the available moves. Each approach has its advantages, and each has led to super-human performance at some games. In this paper, we will look at some major milestones where AI systems first beat humans at checkers, chess, and backgammon; and we will use these systems to discuss some of the strengths and weaknesses of each approach. Then we will look at recent work on playing Go that combines deep neural networks with classical AI methods to achieve super-human performance where neither approach had previously been able to on its own.


Read the full essay here.