Abstract:

Exploring large parameter spaces is a common scenario in modern AI solutions.  Parameter space refers to very high dimensional search spaces formed by combining several parameters. For example, a local storage system with parameters including inode size, journal option, I\/O Scheduler, and dev type can constitute a search space of 24k system configuration. Each of these configurations is a system configuration that has different values of the cost associated with deploying a particular system.

In this work, we explore how deep learning, visual analytics, and evolutionary algorithms can be used to search through very large parameter spaces. Using these exploratory techniques, viable solutions can be presented to the users in many different formats with recommendation systems and data visualization techniques. Recommendation Systems can recommend some solutions which can be explored by humans based on different use cases, on the contrary, visualization techniques with interaction allow users to explore part of the large parameter spaces.

 For Zoom information contact events [at] cs.stonybrook.edu