DeepMind AlphaStar, one of the most interesting examples of Artificial Intelligence

Original article was published on Artificial Intelligence on Medium

DeepMind AlphaStar, one of the most interesting examples of Artificial Intelligence

There are several examples of how artificial intelligence can be applied in the world, from personal assistants to cell phones, even on the new rockets that carry supplies to the ISS.

However, the most notorious example of AI in general application today is AlphaStar.

An artificial intelligence of general application focused on playing increasingly complex strategy games.

StarCraft 2

StarCraft 2

StarCraft 2 is a real-time strategy video game, owned by Blizzard Entertainment.

The worldwide launch was on July 27, 2010.

The game is based on three races, named Terran, Protoss and Zerg. Each of these races has resource collection units, ground and air attack units, and both, as resistance, cost, and speed vary by for each unit.

In the same way, there is a wide variety of maps or battlefields where, using strategy, you can achieve a unit and terrain advantage over the enemy.

The game has 1vs1, 2vs2 and 3vs3 modality.

One of the main elements of StarCraft 2 is the presence of BattleNet. A series of online servers that allow games between players in different locations with a unique ease and user experience.

The first version of Battle Net was released in 1997, being renamed Battle Net 2.0 after the release of SC2.

Machine Learning, Reinforcement Learning, and Video Games

One of the fastest growing applications and examples of artificial intelligence has been in video games.

The beginnings of artificial intelligence in the gaming industry was through supervised learning, where the game were given instructions on how to respond to an action.

The greatest change occurs when the artificial intelligence learning method in made through the use of reward techniques for each correctly performed action of the entity.

An easier way to understand it, would be to compare with educating a child, in case he does something well, he is awarded a reward to continue doing it.In the event that the process or action is not performed correctly, a negative reinforcement is assigned.

In this way, artificial intelligence has the ability to learn by repeating actions.

These actions are classified as good or bad, as well as establishing patterns for their optimal use.

History of AlphaStar

Alphastar vs Liquid

The need to create an example of artificial intelligence based on SC2, arises from the fact that we have the concept of measuring the success of AI in its ability to play games in which humans participate.

After the success that DeepMind had in table games such as Chess or Go, the next step was a video game, and to increase the challenge, a strategy video game.

Building on DeepMind’s technology for reinforcement machine learning, an internal league of StarCraft 2 was created, with the intention that it could play against itself or controlled AI agents.

Thus generating knowledge patterns in good actions, bad actions and generation of strategies.

The challenge of a game like StarCraft is that, much like board games, it is required to be able to visualize possibilities for the future. In the same way, there are millions of different actions combined that can be done to get a result.

AlphaStar’s first games against humans were within the same DeepMind team.

Evolution of AlphaStar

AlphaStar defeated “MaNa”

The next obvious step for this example of artificial intelligence was being able to take on a professional StarCraft 2 player.

On December 12, 2018 AlphaStar played against Dario “TLO” Wünsch of Team Liquid in 1vs1 mode.

The result was surprising given that a human victory was expected, however the result was AlphaStar’s 5–0 victory.

At this point the AI ​​could only play with the Protoss race.

On December 19, 2018 AlphaStar played against Grzegorz “MaNa” Komincz of Team Liquid in 1vs1 mode, again with the Protoss race, MaNa’s specialty.

The result was even more surprising, since because playing with a Protoss professional player, a more even result was expected.

In the end the score was 5–0.

AlphaStar today

I beat AlphaStar

Currently AlphaStar plays in the European Master league of StarCraft 2 in Battle Net, being a milestone among professional players to find a match and try to defeat this example of artificial intelligence.

Unlike artificial intelligence that defeated the two members of Team Liquid, AlphaStar is now increasingly “humanizing”.

This many experts call it a setback, because an artificial intelligence that was invincible at the time, has been deteriorating to have more human features.

An example of this is unit selection speed, map movement, and other traits that humans cannot compete with.

This with the intention that humans can be more competitive against artificial intelligence

Future of Gaming and Artificial Intelligence

It is believe that the advances that have been made within the area of ​​games in artificial intelligence, will soon be reflected in other areas.

These Artificial Intelligence of general application elements can have a major impact on robotics, virtual assistants, and medicine.

“Beyond an entity capable of winning video games, we must see an entity capable of understanding and improving the reasoning, strategy and execution capabilities of a human being.”