Further Examination of Exploration
In this chapter, I examine other approaches to exploration that could be combined with texplore’s model. First, I introduce three domain classes that each suggest a different type of exploration. Then, in Section 6.1, I look at how to perform exploration in domains where a needle-in-a-haystack search is required to find an arbitrarily located reward or transition. In the next section, I look at the opposite case: can we explore better in a domain with a richer, more informative set of state features? Finally, in Section 6.3, I present an algorithm that can learn which of these exploration approaches to adopt on-line, while interacting with the environment. Then I present some empirical comparisons of these approaches against texplore in Section 6.4, before summarizing the chapter in Section 6.5.
KeywordsReward Function Exploration Strategy Random Forest Model Intrinsic Reward External Reward
Unable to display preview. Download preview PDF.