Intelligent Behavior in Autonomous Agents through Optimization and Learning

Abstract

The dream of intelligent robotics performing useful tasks is predicated on the ability to quickly generate complex behaviors. These systems must be able to dynamically react and adapt to unstructured environments and unexpected situations, all without having full awareness of the complexity of the world around them. These behaviors then cannot be pre-planned with prior insight, and indeed many complex tasks might be difficult for humans to even specify algorithmically. A successful method must balance the ability to perform well quickly with prior information while also learning from experience to improve over time; as additional computing resources are increasingly available these ideas are realizable. In this thesis we present work involving the combination of model free and model based paradigms. The interaction of these methods allows for faster behavior synthesis through trajectory optimization while also incorporating experience discovered through planned exploration to enable more long term planning.