Integrated Learning for Goal-Driven Autonomy

Goal-Driven Autonomy (GDA) is an online planning framework that focuses on the integration of planning, execution and goal reasoning. Given a goal, a GDA agent generates a plan to pursue the goal. In addition, by using its expectations, the agent reasons about what the next observed state should be when the plan's actions are executed. If the expectation does not match the observed state, the GDA agent is able to suggest a new goal to be pursued.In most GDA research, knowledge is handcrafted and later fed into the GDA agent by humans who are experts in a particular problem domain. Therefore, in this dissertation, we would like to investigate about how we can create GDA agents that have abilities to acquire knowledge by themselves and reuse that knowledge. The problem domains we focus are real-time strategy (RTS) games. We used two RTS games called DOM game and Wargus. We used Reinforcement Learning because it is an unsupervised learning method and we want our GDA agents to be autonomous. Our research went through multiple steps. First, we built a GDA agent without integration of any learning methods. Later, we incrementally integrated learning methods to each component in the GDA architecture until we build a GDA agent that could learn knowledge for all components. The experimental results show that we can create GDA agents that have the ability to acquire GDA knowledge by themselves.