DeepMind AI researchers have successfully built an AI that has reached GrandMaster ranking in Starcraft 2, beating some of the game's top players.
It's been over 20 years since artificial intelligence began besting players at their own games with the iconic showdown between chess grandmaster Garry Kasparov and IBM's Deep Blue computer. Another major milestone was reached in 2015 when DeepMind Technologies (owned by Google parent company Alphabet) beat professional Go player Lee Sedol with its AlphaGo AI, and later in 2017 when its successor AlphaGo Master beat the world's best Go player Ke Jie.
The complexity of games that AI is mastering has grown significantly in recent years, with new neural network based AI often demonstrating problem solving skills that have been described as almost creative. This week DeepMind added another notch to its belt as a version of it reached GrandMaster ranking in the online video game Starcraft 2.
Starcraft is a realtime strategy game involving complex simulation of war between thousands of units, each with different abilities and stats that makes it much more complex than Go or Chess. Players also have to harvest resources, build their own units to counter what they think the enemy's strategy will be, and plan ahead long-term.
A fog of war mechanic also hides enemy units from sight when out of sight lines from your units, so often players have to predict the actions the enemy will take without visual proof. This is an incredibly hard challenge for an AI to solve, forcing it to make decisions with missing information and send out scouts to gain more sight on enemy forces.
AlphaStar takes this incredibly difficult AI challenge and applies the same kind of model as AlphaGo, training a neural network based on records of past matches played by top human opponents. The matches in Starcraft 2 are a perfect model for this as all actions are recorded so that players can re-watch old games in the game engine and scrutinise their own performance. After training on these matches, AlphaStar was pitted against copies of itself to find weaknesses in its strategy.
The AI has now reached GrandMaster ranking for each of the game's three races, putting it in the top 200 players in its region. There were still many top players that the AI couldn't beat, and those who faced it noticed that it had issues coping with the unexpected if they chose an unorthadox strategy. It was also noted to be lacking on strategic decision-making but better than average on executing its stragies, which may be because it doesn't suffer from the mental load and other issues that players face when controlling large numbers of units.
Source: BBC News
