The ongoing debate about AI and human players in RTs continues. Although it is difficult to make AI competent enough to compete with humans, it is not impossible to create AI that is bad enough. AlphaStar, Stockfish, and Gym-uRTS are some of the most well-known examples of AI-powered RTs. But which is better?
Table of Contents
AlphaStar
The AlphaStar AI has several advantages over human players. First, it is able to interact with the game in a different way than humans do. Human players use their keyboard or mouse to give commands, and they move the camera around to see different parts of the playing area. AlphaStar, on the other hand, can see the entire map at once. This means that the AI does not need to select units to learn what they can and cannot do.
AlphaStar: A Google DeepMind AI has defeated a human pro ten time in StarCraft II. This AI is a technological breakthrough that has demonstrated superior macro- and micro-strategic decision making. While humans may not be able to beat AlphaStar in every game, this AI is a good way to test how well it can compete against human players.
DeepMind: Google’s artificial Intelligence Research Team considered StarCraft the greatest gaming challenge. The game’s strategic nature makes it a natural target for demonstration purposes, but people are still the most skilled players in the sci-fi RTS. The AlphaStar AI is geared toward challenging the best human players at StarCraft II. However, it is unclear if it can actually beat humans in a game, so only time will tell.
In the last January match, AlphaStar’s performance was significantly lower than the human player. This may have been due to insufficient training time, or to additional restrictions such as camera control and APM. This could be an AI issue more generally, which may be related to its goal of increasing sample efficiency. While AlphaStar has made some improvements, humans still have a chance.
Unlike RTS, AlphaStar is already being used in other domains. RTS games aren’t immune to this problem as AI is increasingly being integrated into video games. A good example is the AlphaStar in a crowd-funded YouTube series, AI and Games. The creators of the series are dedicated to exploring artificial intelligence in games and other applications. If you enjoy their work, consider becoming a Patreon supporter.
Stockfish
Recent advances in neural networks make it possible to use neural networks as part of a Chess engine. Unlike the classical rule-based AIs, neural networks are not subject to human error. In fact, many people have made improvements to neural networks using these tools. Stockfish, for example, is an open-source, free chess bot that has been improved over the years by hundreds of developers.
AlphaZero: AlphaZero uses neural network to evaluate positions instead of scanning over 70,000,000 positions per second. Its performance was impressive, and it beat Stockfish in one game. AlphaZero uses custom hardware with four tensor processing units. DeepMind published a research paper that describes how AlphaZero was able to defeat Stockfish in a thousand games and beat the previous world champion.
DeepMind’s AlphaStar League
The success of AlphaStar’s AI is largely due to the training program it used to develop it. DeepMind took the human games it’d watched and trained a neural network from them. These neural networks then forked to create new agents. These agents were then matched against each other in games, and encouraged to choose specialties. This is similar to how humans learn strategy. The goal of this game is to beat human players as much as possible, but in a computer game.
DeepMind trained AlphaStar for 14 days in a StarCraft league with 16 TPUs per agent. The 600 agents who were selected were able to experience the StarCraft equivalent of 200 years of StarCraft playing. AlphaStar was then invited to play against professional players. The players were Dario Wunsch (aka TLO) and Grzegorz Komincz (aka MaNa). They were part of Team Liquid.
DeepMind placed restrictions on the AI during training so that it could perform similar tasks to human players. The AI was limited to performing 22 actions every five second. DeepMind also limited AlphaStar to human-only portions of the map. The team also made sure that the AlphaStar could perform human-level mouse clicks at a speed that matched experienced human players. After 27 days of training, the AI placed in the top 0.5% of European players.
Deep neural networks are the neural network architecture that underpins the artificial intelligence used in training AlphaStar. These neural networks use supervised learning to gather data from human replays and train algorithms to predict the actions of different players. The artificial intelligence then applies these predictions to a large dataset to obtain diverse strategies. This allows it to compete against humans and with humans in a league setting. In addition, the AI is trained by using policies from current agents and a public dataset of anonymized human replays.
AlphaStar beat TLO in the first match using a unique strategy that was largely unknown to humans. Its approach is also unconventional – in the first game, AlphaStar chose to wall off its own base instead of walling off the enemy. In addition, AlphaStar was able to process information about its own base and the positions of other opponents in its team. The AI could accomplish all this without having to split its time between different areas of the map. To make it even more efficient, AlphaStar was able to disable notifications in the game’s settings menu.
Gym-uRTS
After the first wave of cases of coronavirus, gyms began to change their operations. Locker rooms have been shut down, group classes have been suspended, and gyms are implementing intensive cleaning regimens. A gym has an advantage over other venues, however: its membership model allows them to contact its members easily. Some LAB members returned to the gym after the outbreak subsided, despite the temporary ban.