The goal of the project is to create a Deep Reinforcement Learning agent that is capable of playing Shadow Tactics at a competent level. Deep Reinforcement Learning is a field in which neural networks have an agent take actions in an environment and learn from trial and error. A single algorithm is capable of learning and playing a large number of Atari games. The task in this project is difficult because Shadow Tactics is a game that has an extremely large state space and requires long-term planning, which are things that current state-of-the-art reinforcement learning algorithms have trouble with. Our approach uses Hierarchical Deep Reinforcement Learning in order to break the problem down into parts. By using a hierarchy of agents that act on different temporal abstractions we are able to have the agent plan longer, explore better in an environment with sparse reward, and break up the state space into only the relevant parts.
No Team Members