Recommended Free Tools
AlphaStar was a major advance in AI game-playing because it learned to compete in the full, real-time strategy game of StarCraft II—not a simplified, turn-based version. Google DeepMind introduced it in January 2019 after training it first on human match replays, then through reinforcement learning in a league of agents that continually adapted to one another.
That approach helped AlphaStar develop strategies capable of beating professional players and reaching Grandmaster-level performance across all three playable races. Its broader contribution was methodological: it showed how a diverse population of competing agents could help an AI handle hidden information, long-term planning and rapid decisions in a complex environment.
| # | Preview | Product | Price | |
|---|---|---|---|---|
| 1 |
|
StarCraft II: Wings of Liberty | $32.00 | Buy on Amazon |
| 2 |
|
Starcraft II: Legacy of the Void - Standard Edition | $42.94 | Buy on Amazon |
| 3 |
|
Starcraft II PC | $21.45 | Buy on Amazon |
Why StarCraft II was a difficult test for AI
StarCraft II asks players to manage an economy, build a base, scout an opponent, choose a strategy and control combat units—all in real time. Players cannot see everything the opponent is doing, so they must make decisions with incomplete information and anticipate what might happen later.
The game also presents an unusually large set of possible actions. DeepMind’s 2019 account estimated approximately 1026 legal actions at each time-step. That figure describes the scale of the game’s action space, not a number of moves AlphaStar had to execute at once. The challenge was to select useful actions while reacting to a changing match and planning beyond the immediate fight.
#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
Unlike an AI tested on a simplified board or with turns to consider each move, AlphaStar was designed for the full StarCraft II environment. Its task combined strategic planning, imperfect information and real-time control across hundreds of units and buildings.
How AlphaStar learned to play
It began by learning from human replays
AlphaStar’s neural network took in data from the game interface and produced action instructions. Its initial policy was trained by supervised imitation: it learned patterns from anonymized human games. This gave the system a starting point based on strategies people had already developed, rather than requiring it to discover every basic tactic from scratch.
A league of agents then trained against one another
After imitation learning, DeepMind placed agents in a league and used reinforcement learning to improve them through play. Agents could branch into new competitors and adapt to opponents’ tactics. Earlier agents remained in the league, preserving a range of strategies instead of letting training revolve around a single, increasingly familiar opponent.
This diversity mattered because a strategy that succeeds against one rival may fail against another. By repeatedly encountering agents with different approaches, competitors could discover counter-strategies and become less dependent on exploiting one opponent’s habits. DeepMind sampled the final agent from the league’s Nash distribution, a way of representing a mixture of strategies rather than selecting only one specialist.
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 errorsRank #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
The system combined several neural-network components
DeepMind described a transformer over units, a deep LSTM core, an autoregressive policy head, a pointer network and a centralized value baseline. In broad terms, these components helped the system represent the state of many game units, track information over time, select sequences of actions and estimate how promising a situation was. The architecture was built to translate complex game observations into decisions, rather than to rely on a hand-scripted sequence of moves.
Training compressed a vast amount of simulated play
DeepMind reported that the training league ran for 14 days on distributed Google v3 TPUs. During that period, each agent experienced up to 200 years of real-time StarCraft play. These are DeepMind’s reported training figures: the “years” refer to simulated play accumulated by an agent, not the duration of the project or the time a human player would need to reproduce it.
What AlphaStar achieved against humans
DeepMind reported 5–0 results against professional players Grzegorz “MaNa” Komincz and Dario “TLO” Wünsch in its December 2018 evaluation sequence. The results were an important demonstration against elite opponents, but they should be read alongside the later, broader evaluation rather than as a complete description of AlphaStar’s performance.
In a peer-reviewed 2019 study published in Nature, AlphaStar was reported at Grandmaster level for Terran, Zerg and Protoss, and above 99.8% of officially ranked human players. Those findings speak to performance across the ranked player population and all three races; they do not mean AlphaStar was unbeatable or establish a claim about current AI systems.
Free tools Windows power users keep installed
One-click scans. No signup required.
Rank #3
- Videogame Software
| Result | What was reported | Source and date |
|---|---|---|
| Professional evaluation | 5–0 against MaNa and 5–0 against TLO in DeepMind’s reported evaluation sequence | Google DeepMind, 2019, describing the December 19, 2018 benchmark |
| Rank among officially ranked players | Above 99.8% | Nature, 2019 |
| Performance by race | Grandmaster-level ratings for Terran, Zerg and Protoss | Nature, 2019 |
Did AlphaStar have an unfair speed advantage?
DeepMind reported that AlphaStar averaged about 280 actions per minute in its professional games and had an average 350-millisecond delay between observation and action. Those figures help put the results in context: the system’s success was not simply a matter of issuing actions as quickly as possible.
The interface used for evaluation is also important. AlphaStar initially received a raw interface that exposed visible unit attributes without requiring it to move a camera around the map. A later camera-interface version had to choose where to look, adding a decision faced by human players. DeepMind reported that this camera agent exceeded 7000 internal MMR after training. That rating is a reported internal measure and should not be confused with the public ranked-player percentage from the Nature study.
What AlphaStar changed—and what it did not prove
It demonstrated a training approach for strategic opponents
AlphaStar’s league addressed a weakness of training against a single fixed opponent: an agent can become very good at exploiting that opponent without learning strategies that generalize. Keeping diverse competitors in play created opportunities for counter-strategies and encouraged more robust behavior in a multi-agent setting.
It was a milestone in game-playing research, not a general-purpose AI
The reported results apply to StarCraft II and to the evaluation settings described by DeepMind and the 2019 study. Beating professionals in a game does not by itself show that an AI can transfer the same abilities to unrelated tasks, work autonomously in the wider world or reason like a person. AlphaStar’s importance is more specific: it showed how imitation learning and league-based reinforcement learning could be combined to tackle a demanding real-time strategy environment.
Its results are historical
The 2018 professional games and 2019 Grandmaster findings are milestones from that period, not evidence about the strongest AI systems or commercial capabilities available today. The durable lesson is about training design: in a difficult environment with hidden information and many viable strategies, opponents that keep changing can be as important as the agent’s neural-network architecture.
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.
