DeepMind in StarCraft 2 — the humans must be redeemed

Last week DeepMind and Blizzard held a demonstration of AlphaStar, an AI agent playing StarCraft 2. Two professional players Dario "TLO" Wünsch and Grzegorz "MaNa" Komincz participated in the demonstration, playing five games against AlphaStar.

Since AlphaStar defeated the professional gamers in all demonstrated games, many news outlets reported that a machine beat the human in StarCraft 2. What many of them failed to report was that AlphaStar had an unfair advantage, and wouldn’t be successful in a real tournament.

Let me explain how and why.

The machine doesn’t play StarCraft 2 better than human (for now)

At the moment of the demonstration, the AlphaStar agents were playing good enough to defeat many good players, but you can’t call them undefeatable. Here is a video, where Kevin "Hjax" Konrad, a competitive player who develops his own Custom AIs (HjaxBot & Kagamine) explains a simple way to crush AlphaStar.

The simplest way to abuse Alphastar would be to send hallucinations and have the opponent army chase it all over the map. The players were asked to refrain from using hallucinations, but still, a fast flying unit like Oracle or Phoenix could be used for the same purpose. It's easy to get once you analyse a couple of AlphaStar's replays, a courtesy MaNa and TLO were deprived of.

In some games, AlphaStar won by using simple surprise strategies (cheeses) that can’t work more than once in a series. If the same agent played against MaNa or TLO more than one game, the human players would win easily.

The true reason why MaNa and TLO lost all five games

When TLO and MaNa went to play the games, they applied the standard logic for a best of five series: "I need to find out what my opponent does and counter it." What they didn’t know was that DeepMind staff prepared different AlphaStar iterations (agents) for each game. Essentially it means that they played against a different opponent each time. The logic should have been “I need to win five best of ones” instead of “I need to win one best of five series.” The human players didn’t know it.

Another reason why AlphaStar won was that its APM wasn’t capped properly. Good human players average around 300 APM throughout a game, AlphaStar is capped at a similar margin. On the other hand, the machine hasn’t been stolen the ability to burst APM in clutch moments.

It may sound confusing, so here is the explanation: A professional player can burst APM to 500-700 for a couple of seconds during a battle to quickly move units and cast spells. Maintaining 700 APM for 5 seconds results in approximately 60 actions. AlphaStar can expand 250 actions in the same period of time, giving him an unfair advantage over human opponents. After the fight, it will be impaired to one action per second for the next minute, but it’s not very punishing if you are as efficient as a machine.

MaNa isn’t that much stronger than TLO

It’s not that important in the light of the main topic, but I want to restore TLO’s image who has been portrayed as a "weak(er) player."

During the demonstration, TLO played the Protoss vs. Protoss matchup, though his main race is Zerg. It’s as if a Formula 1 pilot was put behind the wheel of a truck — physics work in the same way, you apply similar logic, but many other things are different. The way it’s reported on many sites, makes an impression that TLO is worse than MaNa by a significant margin.

MaNa is probably favored against TLO, but I wouldn’t bet money on him in a fair series where both players use their primary races. Aligulac based on previous results indicates that MaNa would win, but in 2018 TLO kept running into Serral, by all records the best StarCraft 2 player in the world, which makes his stats worse than they really are.