Relational Neurogenesis for Lifelong Learning Agents

Reinforcement learning systems have shown tremendous potential in being able to model meritorious behavior in virtual agents and robots. The ability to learn through continuous reinforcement and interaction with an environment negates the requirement of painstakingly curated datasets and hand crafted features. However, the ability to learn multiple tasks in a sequential manner, referred to as lifelong or continual learning, remains unresolved. The search for lifelong learning algorithms creates the foundation for this work. While there has been much research conducted in supervised learning domains under lifelong learning, the reinforced lifelong learning domain remains open for much exploration. Furthermore, current implementations either concentrate on preserving information in fixed capacity networks, or propose incrementally growing networks which randomly search through an unconstrained solution space. In order to develop a comprehensive lifelong learning algorithm, it seems essential to amalgamate these approaches into a condensed algorithm which can perform both neuroevolution and constrict network growth automatically. This thesis proposes a novel algorithm for continual learning using neurogenesis in reinforcement learning agents. It builds upon existing neuroevolutionary techniques, and incorporates several new mechanisms for limiting the memory resources while expanding neural network learning capacity. The algorithm is tested on a custom set of sequential virtual environments which emulate several meaningful scenarios for intellectually down-scaled species and autonomous robots. Additionally, a library for connecting an unconstrained range of machine learning tools, in a variety of programming languages to the Unity3D simulation engine for the development of future learning algorithms and environments, is also proposed