In the past, video games and games, in general, have provided interesting environments for AI researchers. Recently, neural networks have tackled increasingly complex games, such as Go. Among the top challenges in today's research in that area is the competitively played video game Starcraft 2. From an AI perspective, the game is challenging both in its basic elements (unit control, building placement) as well as in its overall strategy. In particular, the overall strategy, also called Macro gameplay, is interesting due to the need to think ahead, the sheer amount of possible combinations, ability to react to the opponents actions and the creativity that goes into developing these strategies.
In this project, your aim is to write a Starcraft 2 bot. While the basic tasks of the game (building placement) should be automated using scripted agents, the macro gameplay should be learned by using deep neural networks. We plan on having three teams of 3 - 4 students working together on an individual agent.
Proficient programming skills are required to participate in the research practical. Students will need to train themselves in the APIs currently available for the game. They also need to be able to transform abstract descriptions into working algorithms. Lastly, while we do use libraries to simplify some of the programming tasks, we also expect students to learn the theoretic foundations of scripted agents as well as neural networks.
The kick-off meeting will be on the 7th of February at 10:00 in room B233. Participating in this meeting is mandatory.