DeepMind AI goes 10-1 against pro StarCraft 2 players TLO and MaNa

Today, the AI developers at DeepMind showcased AlphaStar, the company’s attempt at bringing machine learning to StarCraft II. The results are not good for us human players of the best strategy games. AlphaStar took down two separate StarCraft pros – TLO and MaNa – with a whopping 10-1 record. The AI’s technique is occasionally wildly off meta, and sometimes just what you’d expect from standard pro play, but either way it’s incredibly effective.

The exhibition broadcast shows the results of two separate series, all Protoss versus Protoss, held over the course of a few weeks. The first five games pitted AlphaStar against TLO, and in those games we see some very non-standard plays from the AI – repeatedly pushing up ramps, failing to wall off its base early on, building way more probes than usual, and favouring unusual army configurations like tonnes of Disruptors.

While none of that’s standard play for Protoss mirrors, AlphaStar still went 5-0 in that first set. TLO is a Zerg main, so the next round was more of a like versus like matchup against Protoss player MaNa. The AI got an extra week of training in the meantime, and it played a bit different – but still won 5-0 in a more challenging match.

Those were largely competitive matches, but the AI continued to win in the end, in large part due to stronger unit control. In an interview during the stream MaNa also mentioned AlphaStar’s unusual strategies – they’re things not typically seen in high-level human play, so they’re tough to predict and counter. Even then, by the end the AlphaStar strats start to look more and more human.

AlphaStar’s one loss came from a live exhibition match that took place during the stream. The AI had an early lead, but made some mistakes in mid-game exchanges that eventually led to the single blemish on its record.

The AI trains essentially by watching high level replays, branching off, and competing against itself, and a week of training sees an entire league of AlphaStar ‘agents’ compete against one another while refining their strategies. For each of these exhibition sets, the best AI branches were selected to compete against the pros.

You can watch the archive of the exhibition above, including plenty more detail about how AlphaStar works and what its limitations are. For AI researchers, SC2 is a particularly interesting challenge, since unlike Chess and Go players don’t know the state of the entire map and have to make their actions in real time. That means our robot overlords are one step closer to total world domination. Make your preparations accordingly.