A Toolkit for Encouraging Safe Diversity in Skill Discovery

Diversifying agents' skills has proven to be critical for adapting to a wide range of tasks. However, continuously promoting diversity can have catastrophic effects, such as accumulating unsafe, ineffective or misaligned skills. To avoid such outcomes, providing agents with the ability to modulate diversity in skill discovery remains a largely unexplored research area. In my research, I aim to design agents that can control and adapt their diversity to fit any context. Integrating context into skill discovery was my initial approach to controlling diversity. This was done by allowing the agent to use human preferences to identify regions of the environment where diversity is most likely to be desired. However, to modulate skill diversity, an agent has to be able to not only identify where to demonstrate diversity, but also comprehend how it affects the environment and others around it to decide when to be more (or less) diverse. To achieve this, an agent needs more tools, such as observing its own diversity and methods of adjusting it. The incorporation of controlled diversity will, I believe, make agents with multiple behaviors more flexible, reliable, and robustly applicable in a wide variety of contexts.