In this talk, I will discuss the idea of having agents learn a library of behaviors for solving sequential-decision-making problems. Similarly to how programmers develop libraries of reusable code, I will argue that artificial agents should be equipped with similar abilities. Given a stream of tasks, the agent will continually learn different behaviors that are stored in a library; the agent will also learn how to use such a library of behaviors to solve new incoming tasks. This library-driven learning approach fosters the reusability of learned concepts through the composition of existing behaviors. I will show examples of agents that maintain a library of behaviors to speed up the process of learning policies. In these examples, the agent learns by maintaining a library of programs derived from searching in the space of programs a programming language defines, or from the decomposition of neural networks into sub-networks encoding reusable behaviors.
Back to Workshop III: Naturalistic Approaches to Artificial Intelligence
Copyright. All Rights Reserved.
