Empirical Domains

Abstract

In this chapter we introduce the testbed domains used in the remainder of this monograph, where we informally define a domain as a setting for one or more tasks (i.e., MDPs). The primary purpose of this chapter is to provide sufficient background to allow a reader to fully understand the transfer learning experiments in subsequent chapters. In addition to describing each domain, we explain how each can be learned with one or more of the RL methods discussed in the previous chapter. In every domain we will emphasize how some tasks are faster to master than others. In general, experiments in this monograph transfer from relatively simple, quick to learn, source task to a more complex target task. In the target task time scenario, an effective TL algorithm may reduce the target task learning time regardless of task ordering, but only by ordering tasks in order of increasing difficulty will the total training time be reduced.