Deep Reinforcement Learning in Starcraft 2

Last year DeepMind and Blizzard released a set of APIs designed to act as a test-bed for learning algorithms. After the success of systems like AlphaGo in defeating the Go world champion, the AI community has pivoted to working on even more complex games like Starcraft II and Dota 2. The theory being that if an AI system can learn to play complex computer games it will eventually be able to develop some sort of general intelligence. So don't be surprised if in the next few years you read headlines like "Menacing Super-AI Defeats Starcraft World Champion".

Last semester I worked on the problem as a part of a class project. I didn't achieve anything truly spectacular, but did see some evidence of learning. The learning algorithms create an agent that goes from brain-dead to actually using some pretty respectable strategies--all without any human input. Notice how the agent sends out marines as pawns to bait the enemy into attacking? The next 20-30 years of AI research are going to be exciting if the current pace keeps up.