This master's thesis documents the work of applying reinforcement learning on various subtasks in a commercial-quality RTS game.

Some of the main problems when working with reinforcement learning are convergence rates, handling concurrent agents, and minimising the state space. Some ideas to solve these problems, are to include time as an important factor during learning, and to decompose the state space by identifying independent objects in the game world. Regarding concurrent agents, we investigate how much information each agent needs about the other agents in order to behave optimally.

We found that, given certain restrictions on the scenarios, we are able to improve convergence rate as well as minimising the state space for a task. We also identify how much concurrency information is suitable for concurrent agents in various scenarios. Finally, we provide a short discussion of how we can combine our solutions to solve even more complex problems.