Reinforcement learning applied to games
- PDF / 2,508,280 Bytes
- 16 Pages / 595.276 x 790.866 pts Page_size
- 17 Downloads / 289 Views
Reinforcement learning applied to games João Crespo1 · Andreas Wichert1 Received: 23 November 2019 / Accepted: 17 March 2020 / Published online: 6 April 2020 © Springer Nature Switzerland AG 2020
Abstract With the continuous evolution of computational power, especially in the computer graphics area, reinforcement learning has been gaining traction in the community as many novel methods are being created and older ones revamped. Many of these employ artificial neural networks, yet treat them as a black box system. In this paper, we start by introducing classic and recent developments in the area of machine learning, followed by an overview of the most relevant work done with deep neural networks applied to games. Our objective is to give the reader insights on how neural networks learn and clarify the decisions made in the development of modern systems. Keywords Reinforcement learning · Neural network · Games
1 Introduction Artificial intelligence (AI) systems have been a part of computer science for many years now and have been helping humans with multiple tasks ranging from medical diagnosis to self-driving cars. Most of the systems developed in the first decades of AI were algorithmic and made up of rules based on human expertise applied to algorithms such as minimax and alpha–beta pruning [16]. However, the goal has always been to develop an agent that could analyze, learn and act upon the environment on its own. Machine learning is the area of artificial intelligence concerned with the development of models that use data or experience to learn how to succeed in a certain task. For instance, in [8] neural networks were used to analyze eye scans and detect signs of eye disease with a better accuracy than experts in the field. Moreover, in [49] an agent was trained to produce music based on previous examples made by humans. Games provide proof of concept in a sandbox-like environment that allows us to experiment and develop key ideas that might be useful in other areas. For instance, DeepMind used key aspects they learned while developing AlphaZero [36, 37], an agent capable
of playing chess, go or shogi, to project AlphaFold’s [27] neural network architecture. This system is capable of predicting the 3D shape of proteins, which can help scientists understand how to treat diseases they believe are caused by badly shaped proteins. In this paper, we focus mainly in one type of machine learning called reinforcement learning and how it has been used to play various games. Firstly, we introduce some basic reinforcement learning concepts and algorithms designed throughout the past years. Secondly, we explore practical implementations of these algorithms and analyze some novel proposed algorithms that make use of the concepts from the first part, like AlphaGo [35], the first engine to beat the world champion in the game of Go, and proximal policy optimization [34]. With this work, we hope to gain better insights on how reinforcement learning systems work and how we can improve them. The rest of the paper is organiz
Data Loading...
