Deepmind – a Google-owned artificial intelligence company – is very well-known for their creations. Their “children” have a history of winning over actual human. Not any ordinary human, some “children” of Deepmind have encountered and won over world-class players in their field. In 2017, AlphaGo by Deepmind took on the Go world champion Ke Jie and the victory belonged to AlphaGo, which made AlphaGo the best Go player on Earth, although it is not a human. Recently, Deepmind and Google stepped their next move into esports field by introducing their new child – AlphaStar. This AI agent is designed to play the game StarCraft II. The first testers of AlphaStar were 2 StarCraft II pro players: TLO and MaNa. And what happened after that shocked us all: AlphaStar won 10 games in a row before TLO and MaNa were able to pick up one last win. This last match between AlphaStar and MaNa was live-streamed by both Blizzard and Deepmind, and until this match MaNa was able to beat AlphaStar, after ten losses continuously.

The map used for the whole competition was Catalyst and the game version for the competition was slightly old (AlphaStar could only run that version in the time of the competition). TLO, on his stream, declared that he would beat AlphaStar, but the result was a 5-0 victory for AlphaStar over TLO. What even more surprising is that AlphaStar managed to use five different strategies for five games

Of course, we have to acknowledge that AlphaStar had some advantages in the matchup with TLO. First, the match used Protoss as the class of units for all five games (Protoss is not TLO main class). Second, basically, AlphaStar sees the map entirely zoomed out. Although AlphaStar's vision was still blocked by the fog of war, it can process any visible information at once without having to divide its time to process information in different part of the map. This gave AlphaStar a massive opportunity against TLO as well as any other human players since we all know how much time does it normally take us to rapidly move the screen to control the whole map.

1334
The way AlphaStar sees the map

This enabled AlphaStar to make smarter and more efficient decisions, which might explain the poor statistics of AlphaStar. In both the match with TLO and the match in MaNa later on, AlphaStar actually did perform fewer actions per minute than its competitors. This statistics from AlphaStar is significantly fewer than the average pro player’s. Moreover, the reaction time of AlphaStar measured in 2 matches was about 0.35s (350 milliseconds), which is slower than most pro players. But the fact that AlphaStar was able to top both TLO and MaNa proved that you don’t have to make many actions or have quick response to win. You only have to make smarter and more efficient actions on the whole picture to win.

45363647

To prepare for the matches, Deepmind designed an in-depth training program for AlphaStar. In the AlphaStar League (the name of the training program) AlphaStar has gone through a lot of human replays and training on them. After one week of intense training, AlphaStar benefited a little bit from each match and was able to form its own strategy. Estimated, after that one week, AlphaStar has acquired the knowledge of StarCraft II that equal to 200 years worth of StarCraft II. After training, AlphaStar took on TLO, and it chose five strategies that least likely to be exploited. The result, as we all knew, was a 5-0 swept from AlphaStar against TLO.

Seeing that AlphaStar can defeat an off-race pro player, Deedmind put on a Protoss pro player. MaNa, a two-time champion of major StarCraft II tournaments and a Protoss main, was chosen. Before the match with MaNa, AlphaStar had another training in a week. This training was more sophisticated than the previous one, and it included the knowledge from 5 games against TLO. After the second training, the improvement of AlphaStar was so noticeable that commentators said it played so like a human. It made less silly mistakes and gave out effective decisions in style. In the second match with MaNa, despite MaNa’s effort in every game, AlphaStar once again swept MaNa with a 5-0 victory, creating a 10-0 record of itself in its very first times facing world-class human’s intelligence in their field.

After the 10-0 victory of AlphaStar, Deepmind introduced a new version of AlphaStar. This version also got the training from the knowledge of the five games with MaNa. However, this version of AlphaStar cannot see the map like its previous version. This new version has to see the map like a normal human being, not see the map as a zoom out. This made the AI agent has to act, play and place its focus the same way as a human player. After the training for this version of AlphaStar, MaNa had his second chance when they faced each other in a live-streamed match by Blizzard and DeepMind.

This time, MaNa did pull out his victory due to the mistakes and shortcomings by AlphaStar. These mistakes looked like they came from the view restriction of this new AlphaStar version.

Of course, it is really hard for us to truly determine the reason for the first 10-0 loss of TLO and MaNa. Is it because of the unrestricted view of AlphaStar? Or is it the fact that AI is getting better than a human’s mind? No matter what the true answer is, this event is definitely not good news for pro gamers. Especially a two-time champion of major StarCraft II tournaments got beaten 5-0 in his preferred race. While pro gamers are certainly concerning, regular gamers like us can definitely benefit from this. Anyone can easily access the replays and the analysis of the games between TLO and AlphaStar, as well as the game between MaNa and AlphaStar. Deepmind has posted them on their website. Please keep in mind that this is an AI agent that has gone through more than 200 years of StarCraft II training, so you will definitely learn something (actually a lot) from the replays and the deep analysis. About TLO and MaNa, as well as all other pro players, perhaps it’s time for them to level up their gameplay.