Hierarchical Reinforcement Learning and Safe State Abstraction on Realtime Strategy Games
Folkert Huizinga, Supervised B. Bakker · 2007
For an agent to perform well in a dynamic environment, such as those in computer games, adaptation is required. However, the agents in the current games are scripted, which means they will not learn from experience and therefore do not adapt. Because of this lack of learning and adapting, it makes the current agents boring to play against as a human being. This paper presents a non scripted, non reactive approach to a simplified real time strategy game, using Hierarchical Reinforcement Learning, in particular the MaxQ method with safe state abstraction [1]. We will compare it to a Qlearning algorithm and MaxQ without safe state abstraction.