Quick wins for a faster PC:
Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Repair Windows errors before they cause bigger problemsFix Now →Scan for outdated or missing drivers - takes under a minuteDriver Scan →DeepMind’s AlphaStar Final reached Grandmaster level with all three StarCraft II races and was rated above 99.8% of officially ranked human players in an October 2019 evaluation. That figure is a ladder-rating percentile—not a record of beating 99.8% of people in direct matches. The comparison covered roughly 90,000 recently active, league-eligible players on the European server.
How did DeepMind’s AlphaStar beat 99.8% of StarCraft II players?
AlphaStar Final earned Match Making Ratings (MMR) of 6,275 as Protoss, 6,048 as Terran, and 5,835 as Zerg. The peer-reviewed Nature paper reports that its rating was above 99.8% of officially ranked players active enough in recent months to enter a league—approximately 90,000 players on the European server.
In other words, the result places AlphaStar near the top of the evaluated ladder population. It does not mean the system played every human, defeated 99.8% of them in head-to-head games, or could not be beaten. The result is specific to the 2019 experiment and its game version, server, matchmaking, and play constraints.
What was tested, and how?
The evaluation took place online through Battle.net using StarCraft II balance patch 4.9.3. Players opted into the experiment, and their accounts were anonymous. The study used four maps and compared AlphaStar’s matchmaking ratings with those of eligible human players.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Fix the driver behind crashes, sound loss and screen glitches3Clear out junk files and repair common Windows errors#1 Best Overall
- Fast-paced, hard-hitting, tightly balanced competitive real-time strategy gameplay that recaptures and improves on the original game
- Three completely distinct races: Protoss, Terran, and Zerg
- Units and gameplay mechanics distinguish each race
- 3D-graphics engine with support for visual effects and massive unit and army sizes
- Full multiplayer support, with competitive features and matchmaking utilities available through Battle.net
In the final evaluation phase, AlphaStar played 30 games per race, starting from the midpoint rating. Earlier evaluation stages used 30 or 60 games per race. DeepMind’s account describes 44 days of League training before AlphaStar Final; these figures and ratings are historical results, not current ladder standings.
AlphaStar played the full game through a camera interface, with restrictions on its action rate. DeepMind’s account of the system gives a maximum of 22 agent actions per five seconds. An agent action could count as up to three in-game actions per minute (APM) actions, and camera movement also counted as an agent action. These were designed constraints, not proof that the system used an ordinary human setup in every detail.
How did AlphaStar learn to play?
AlphaStar combined learning from human games with reinforcement learning and self-play. Rather than train a single agent against a fixed opponent, DeepMind developed a League of agents that continually generated strategies and counter-strategies.
Rank #2
- This is a standalone product. It does not require any other version of StarCraft II to play
- Internet Connection Required
- Battle.net registration and Battle.net Desktop Application required
Human-game imitation
Supervised learning from human games gave agents examples of how people play. This imitation-learning stage helped establish useful behavior before and alongside reinforcement learning.
The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Reinforcement learning and self-play
Agents improved by playing games and learning from their outcomes. Self-play let them test and refine strategies against other agents rather than relying only on human examples.
A League that searched for weaknesses
The League included main agents pursuing broad performance and exploiter agents designed to find weaknesses in other strategies. The system also used imitation learning and distillation to help preserve useful behavior and reduce forgetting as agents adapted. The Nature paper describes training on human and agent games within this changing ecosystem of strategies and counter-strategies.
What the 99.8% result does—and does not—show
The study’s blind matchmaking was chosen to estimate AlphaStar’s performance under approximately stationary conditions. The authors explicitly say this setup does not directly measure how susceptible the agent would be to exploitation under repeated play. A high rating in this evaluation therefore demonstrates exceptional ladder strength under the tested conditions, not invulnerability to an opponent who repeatedly studies and targets its behavior.
The paper also notes that some cross-race matchups had limited per-race data. That qualification matters when interpreting individual race results: the headline percentile summarizes the evaluation, but it should not be mistaken for an equally extensive test of every possible matchup or strategy.
Free tools Windows power users keep installed
One-click scans. No signup required.
Why AlphaStar was a milestone for game-playing AI
StarCraft II requires an agent to make decisions under uncertainty, manage resources, build an economy and army, and respond to an opponent’s changing plans. AlphaStar’s result showed that a system trained through human examples, reinforcement learning, self-play, and a diverse League could reach Grandmaster-level ratings across Protoss, Terran, and Zerg—not just perform well with one race or in a narrow scripted scenario.
Rank #4
- Used Book in Good Condition
Professional player Dario “TLO” Wünsch described the system as strategically skilled at judging when to engage or disengage, while saying its play did not feel superhuman or beyond what a human could theoretically achieve. That is a player’s impression, not a separate measurement of the experiment’s performance.
Sources and scope
- Nature (2019): “Grandmaster level in StarCraft II using multi-agent reinforcement learning” — peer-reviewed study and evaluation results.
- Google DeepMind (30 October 2019): “AlphaStar: Grandmaster level in StarCraft II using multi-agent reinforcement learning” — project account and interface/action-rate details.
AlphaStar was a research system described in these sources, not a consumer AI product. The reported standings refer to the 2019 evaluation; they should not be read as current rankings.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
